Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * linux/mm/slab.c | |
3 | * Written by Mark Hemment, 1996/97. | |
4 | * (markhe@nextd.demon.co.uk) | |
5 | * | |
6 | * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli | |
7 | * | |
8 | * Major cleanup, different bufctl logic, per-cpu arrays | |
9 | * (c) 2000 Manfred Spraul | |
10 | * | |
11 | * Cleanup, make the head arrays unconditional, preparation for NUMA | |
12 | * (c) 2002 Manfred Spraul | |
13 | * | |
14 | * An implementation of the Slab Allocator as described in outline in; | |
15 | * UNIX Internals: The New Frontiers by Uresh Vahalia | |
16 | * Pub: Prentice Hall ISBN 0-13-101908-2 | |
17 | * or with a little more detail in; | |
18 | * The Slab Allocator: An Object-Caching Kernel Memory Allocator | |
19 | * Jeff Bonwick (Sun Microsystems). | |
20 | * Presented at: USENIX Summer 1994 Technical Conference | |
21 | * | |
22 | * The memory is organized in caches, one cache for each object type. | |
23 | * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct) | |
24 | * Each cache consists out of many slabs (they are small (usually one | |
25 | * page long) and always contiguous), and each slab contains multiple | |
26 | * initialized objects. | |
27 | * | |
28 | * This means, that your constructor is used only for newly allocated | |
29 | * slabs and you must pass objects with the same intializations to | |
30 | * kmem_cache_free. | |
31 | * | |
32 | * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM, | |
33 | * normal). If you need a special memory type, then must create a new | |
34 | * cache for that memory type. | |
35 | * | |
36 | * In order to reduce fragmentation, the slabs are sorted in 3 groups: | |
37 | * full slabs with 0 free objects | |
38 | * partial slabs | |
39 | * empty slabs with no allocated objects | |
40 | * | |
41 | * If partial slabs exist, then new allocations come from these slabs, | |
42 | * otherwise from empty slabs or new slabs are allocated. | |
43 | * | |
44 | * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache | |
45 | * during kmem_cache_destroy(). The caller must prevent concurrent allocs. | |
46 | * | |
47 | * Each cache has a short per-cpu head array, most allocs | |
48 | * and frees go into that array, and if that array overflows, then 1/2 | |
49 | * of the entries in the array are given back into the global cache. | |
50 | * The head array is strictly LIFO and should improve the cache hit rates. | |
51 | * On SMP, it additionally reduces the spinlock operations. | |
52 | * | |
53 | * The c_cpuarray may not be read with enabled local interrupts - | |
54 | * it's changed with a smp_call_function(). | |
55 | * | |
56 | * SMP synchronization: | |
57 | * constructors and destructors are called without any locking. | |
58 | * Several members in kmem_cache_t and struct slab never change, they | |
59 | * are accessed without any locking. | |
60 | * The per-cpu arrays are never accessed from the wrong cpu, no locking, | |
61 | * and local interrupts are disabled so slab code is preempt-safe. | |
62 | * The non-constant members are protected with a per-cache irq spinlock. | |
63 | * | |
64 | * Many thanks to Mark Hemment, who wrote another per-cpu slab patch | |
65 | * in 2000 - many ideas in the current implementation are derived from | |
66 | * his patch. | |
67 | * | |
68 | * Further notes from the original documentation: | |
69 | * | |
70 | * 11 April '97. Started multi-threading - markhe | |
71 | * The global cache-chain is protected by the semaphore 'cache_chain_sem'. | |
72 | * The sem is only needed when accessing/extending the cache-chain, which | |
73 | * can never happen inside an interrupt (kmem_cache_create(), | |
74 | * kmem_cache_shrink() and kmem_cache_reap()). | |
75 | * | |
76 | * At present, each engine can be growing a cache. This should be blocked. | |
77 | * | |
e498be7d CL |
78 | * 15 March 2005. NUMA slab allocator. |
79 | * Shai Fultheim <shai@scalex86.org>. | |
80 | * Shobhit Dayal <shobhit@calsoftinc.com> | |
81 | * Alok N Kataria <alokk@calsoftinc.com> | |
82 | * Christoph Lameter <christoph@lameter.com> | |
83 | * | |
84 | * Modified the slab allocator to be node aware on NUMA systems. | |
85 | * Each node has its own list of partial, free and full slabs. | |
86 | * All object allocations for a node occur from node specific slab lists. | |
1da177e4 LT |
87 | */ |
88 | ||
89 | #include <linux/config.h> | |
90 | #include <linux/slab.h> | |
91 | #include <linux/mm.h> | |
92 | #include <linux/swap.h> | |
93 | #include <linux/cache.h> | |
94 | #include <linux/interrupt.h> | |
95 | #include <linux/init.h> | |
96 | #include <linux/compiler.h> | |
97 | #include <linux/seq_file.h> | |
98 | #include <linux/notifier.h> | |
99 | #include <linux/kallsyms.h> | |
100 | #include <linux/cpu.h> | |
101 | #include <linux/sysctl.h> | |
102 | #include <linux/module.h> | |
103 | #include <linux/rcupdate.h> | |
543537bd | 104 | #include <linux/string.h> |
e498be7d | 105 | #include <linux/nodemask.h> |
1da177e4 LT |
106 | |
107 | #include <asm/uaccess.h> | |
108 | #include <asm/cacheflush.h> | |
109 | #include <asm/tlbflush.h> | |
110 | #include <asm/page.h> | |
111 | ||
112 | /* | |
113 | * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL, | |
114 | * SLAB_RED_ZONE & SLAB_POISON. | |
115 | * 0 for faster, smaller code (especially in the critical paths). | |
116 | * | |
117 | * STATS - 1 to collect stats for /proc/slabinfo. | |
118 | * 0 for faster, smaller code (especially in the critical paths). | |
119 | * | |
120 | * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible) | |
121 | */ | |
122 | ||
123 | #ifdef CONFIG_DEBUG_SLAB | |
124 | #define DEBUG 1 | |
125 | #define STATS 1 | |
126 | #define FORCED_DEBUG 1 | |
127 | #else | |
128 | #define DEBUG 0 | |
129 | #define STATS 0 | |
130 | #define FORCED_DEBUG 0 | |
131 | #endif | |
132 | ||
1da177e4 LT |
133 | /* Shouldn't this be in a header file somewhere? */ |
134 | #define BYTES_PER_WORD sizeof(void *) | |
135 | ||
136 | #ifndef cache_line_size | |
137 | #define cache_line_size() L1_CACHE_BYTES | |
138 | #endif | |
139 | ||
140 | #ifndef ARCH_KMALLOC_MINALIGN | |
141 | /* | |
142 | * Enforce a minimum alignment for the kmalloc caches. | |
143 | * Usually, the kmalloc caches are cache_line_size() aligned, except when | |
144 | * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned. | |
145 | * Some archs want to perform DMA into kmalloc caches and need a guaranteed | |
146 | * alignment larger than BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that. | |
147 | * Note that this flag disables some debug features. | |
148 | */ | |
149 | #define ARCH_KMALLOC_MINALIGN 0 | |
150 | #endif | |
151 | ||
152 | #ifndef ARCH_SLAB_MINALIGN | |
153 | /* | |
154 | * Enforce a minimum alignment for all caches. | |
155 | * Intended for archs that get misalignment faults even for BYTES_PER_WORD | |
156 | * aligned buffers. Includes ARCH_KMALLOC_MINALIGN. | |
157 | * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables | |
158 | * some debug features. | |
159 | */ | |
160 | #define ARCH_SLAB_MINALIGN 0 | |
161 | #endif | |
162 | ||
163 | #ifndef ARCH_KMALLOC_FLAGS | |
164 | #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN | |
165 | #endif | |
166 | ||
167 | /* Legal flag mask for kmem_cache_create(). */ | |
168 | #if DEBUG | |
169 | # define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \ | |
170 | SLAB_POISON | SLAB_HWCACHE_ALIGN | \ | |
171 | SLAB_NO_REAP | SLAB_CACHE_DMA | \ | |
172 | SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \ | |
173 | SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ | |
174 | SLAB_DESTROY_BY_RCU) | |
175 | #else | |
176 | # define CREATE_MASK (SLAB_HWCACHE_ALIGN | SLAB_NO_REAP | \ | |
177 | SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \ | |
178 | SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ | |
179 | SLAB_DESTROY_BY_RCU) | |
180 | #endif | |
181 | ||
182 | /* | |
183 | * kmem_bufctl_t: | |
184 | * | |
185 | * Bufctl's are used for linking objs within a slab | |
186 | * linked offsets. | |
187 | * | |
188 | * This implementation relies on "struct page" for locating the cache & | |
189 | * slab an object belongs to. | |
190 | * This allows the bufctl structure to be small (one int), but limits | |
191 | * the number of objects a slab (not a cache) can contain when off-slab | |
192 | * bufctls are used. The limit is the size of the largest general cache | |
193 | * that does not use off-slab slabs. | |
194 | * For 32bit archs with 4 kB pages, is this 56. | |
195 | * This is not serious, as it is only for large objects, when it is unwise | |
196 | * to have too many per slab. | |
197 | * Note: This limit can be raised by introducing a general cache whose size | |
198 | * is less than 512 (PAGE_SIZE<<3), but greater than 256. | |
199 | */ | |
200 | ||
fa5b08d5 | 201 | typedef unsigned int kmem_bufctl_t; |
1da177e4 LT |
202 | #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0) |
203 | #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1) | |
204 | #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-2) | |
205 | ||
206 | /* Max number of objs-per-slab for caches which use off-slab slabs. | |
207 | * Needed to avoid a possible looping condition in cache_grow(). | |
208 | */ | |
209 | static unsigned long offslab_limit; | |
210 | ||
211 | /* | |
212 | * struct slab | |
213 | * | |
214 | * Manages the objs in a slab. Placed either at the beginning of mem allocated | |
215 | * for a slab, or allocated from an general cache. | |
216 | * Slabs are chained into three list: fully used, partial, fully free slabs. | |
217 | */ | |
218 | struct slab { | |
b28a02de PE |
219 | struct list_head list; |
220 | unsigned long colouroff; | |
221 | void *s_mem; /* including colour offset */ | |
222 | unsigned int inuse; /* num of objs active in slab */ | |
223 | kmem_bufctl_t free; | |
224 | unsigned short nodeid; | |
1da177e4 LT |
225 | }; |
226 | ||
227 | /* | |
228 | * struct slab_rcu | |
229 | * | |
230 | * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to | |
231 | * arrange for kmem_freepages to be called via RCU. This is useful if | |
232 | * we need to approach a kernel structure obliquely, from its address | |
233 | * obtained without the usual locking. We can lock the structure to | |
234 | * stabilize it and check it's still at the given address, only if we | |
235 | * can be sure that the memory has not been meanwhile reused for some | |
236 | * other kind of object (which our subsystem's lock might corrupt). | |
237 | * | |
238 | * rcu_read_lock before reading the address, then rcu_read_unlock after | |
239 | * taking the spinlock within the structure expected at that address. | |
240 | * | |
241 | * We assume struct slab_rcu can overlay struct slab when destroying. | |
242 | */ | |
243 | struct slab_rcu { | |
b28a02de PE |
244 | struct rcu_head head; |
245 | kmem_cache_t *cachep; | |
246 | void *addr; | |
1da177e4 LT |
247 | }; |
248 | ||
249 | /* | |
250 | * struct array_cache | |
251 | * | |
1da177e4 LT |
252 | * Purpose: |
253 | * - LIFO ordering, to hand out cache-warm objects from _alloc | |
254 | * - reduce the number of linked list operations | |
255 | * - reduce spinlock operations | |
256 | * | |
257 | * The limit is stored in the per-cpu structure to reduce the data cache | |
258 | * footprint. | |
259 | * | |
260 | */ | |
261 | struct array_cache { | |
262 | unsigned int avail; | |
263 | unsigned int limit; | |
264 | unsigned int batchcount; | |
265 | unsigned int touched; | |
e498be7d CL |
266 | spinlock_t lock; |
267 | void *entry[0]; /* | |
268 | * Must have this definition in here for the proper | |
269 | * alignment of array_cache. Also simplifies accessing | |
270 | * the entries. | |
271 | * [0] is for gcc 2.95. It should really be []. | |
272 | */ | |
1da177e4 LT |
273 | }; |
274 | ||
275 | /* bootstrap: The caches do not work without cpuarrays anymore, | |
276 | * but the cpuarrays are allocated from the generic caches... | |
277 | */ | |
278 | #define BOOT_CPUCACHE_ENTRIES 1 | |
279 | struct arraycache_init { | |
280 | struct array_cache cache; | |
b28a02de | 281 | void *entries[BOOT_CPUCACHE_ENTRIES]; |
1da177e4 LT |
282 | }; |
283 | ||
284 | /* | |
e498be7d | 285 | * The slab lists for all objects. |
1da177e4 LT |
286 | */ |
287 | struct kmem_list3 { | |
b28a02de PE |
288 | struct list_head slabs_partial; /* partial list first, better asm code */ |
289 | struct list_head slabs_full; | |
290 | struct list_head slabs_free; | |
291 | unsigned long free_objects; | |
292 | unsigned long next_reap; | |
293 | int free_touched; | |
294 | unsigned int free_limit; | |
295 | spinlock_t list_lock; | |
296 | struct array_cache *shared; /* shared per node */ | |
297 | struct array_cache **alien; /* on other nodes */ | |
1da177e4 LT |
298 | }; |
299 | ||
e498be7d CL |
300 | /* |
301 | * Need this for bootstrapping a per node allocator. | |
302 | */ | |
303 | #define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1) | |
304 | struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS]; | |
305 | #define CACHE_CACHE 0 | |
306 | #define SIZE_AC 1 | |
307 | #define SIZE_L3 (1 + MAX_NUMNODES) | |
308 | ||
309 | /* | |
7243cc05 | 310 | * This function must be completely optimized away if |
e498be7d CL |
311 | * a constant is passed to it. Mostly the same as |
312 | * what is in linux/slab.h except it returns an | |
313 | * index. | |
314 | */ | |
7243cc05 | 315 | static __always_inline int index_of(const size_t size) |
e498be7d CL |
316 | { |
317 | if (__builtin_constant_p(size)) { | |
318 | int i = 0; | |
319 | ||
320 | #define CACHE(x) \ | |
321 | if (size <=x) \ | |
322 | return i; \ | |
323 | else \ | |
324 | i++; | |
325 | #include "linux/kmalloc_sizes.h" | |
326 | #undef CACHE | |
327 | { | |
328 | extern void __bad_size(void); | |
329 | __bad_size(); | |
330 | } | |
7243cc05 IK |
331 | } else |
332 | BUG(); | |
e498be7d CL |
333 | return 0; |
334 | } | |
335 | ||
336 | #define INDEX_AC index_of(sizeof(struct arraycache_init)) | |
337 | #define INDEX_L3 index_of(sizeof(struct kmem_list3)) | |
1da177e4 | 338 | |
e498be7d CL |
339 | static inline void kmem_list3_init(struct kmem_list3 *parent) |
340 | { | |
341 | INIT_LIST_HEAD(&parent->slabs_full); | |
342 | INIT_LIST_HEAD(&parent->slabs_partial); | |
343 | INIT_LIST_HEAD(&parent->slabs_free); | |
344 | parent->shared = NULL; | |
345 | parent->alien = NULL; | |
346 | spin_lock_init(&parent->list_lock); | |
347 | parent->free_objects = 0; | |
348 | parent->free_touched = 0; | |
349 | } | |
350 | ||
351 | #define MAKE_LIST(cachep, listp, slab, nodeid) \ | |
352 | do { \ | |
353 | INIT_LIST_HEAD(listp); \ | |
354 | list_splice(&(cachep->nodelists[nodeid]->slab), listp); \ | |
355 | } while (0) | |
356 | ||
357 | #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \ | |
358 | do { \ | |
359 | MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \ | |
360 | MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \ | |
361 | MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \ | |
362 | } while (0) | |
1da177e4 LT |
363 | |
364 | /* | |
365 | * kmem_cache_t | |
366 | * | |
367 | * manages a cache. | |
368 | */ | |
b28a02de | 369 | |
2109a2d1 | 370 | struct kmem_cache { |
1da177e4 | 371 | /* 1) per-cpu data, touched during every alloc/free */ |
b28a02de PE |
372 | struct array_cache *array[NR_CPUS]; |
373 | unsigned int batchcount; | |
374 | unsigned int limit; | |
375 | unsigned int shared; | |
376 | unsigned int objsize; | |
e498be7d | 377 | /* 2) touched by every alloc & free from the backend */ |
b28a02de PE |
378 | struct kmem_list3 *nodelists[MAX_NUMNODES]; |
379 | unsigned int flags; /* constant flags */ | |
380 | unsigned int num; /* # of objs per slab */ | |
381 | spinlock_t spinlock; | |
1da177e4 LT |
382 | |
383 | /* 3) cache_grow/shrink */ | |
384 | /* order of pgs per slab (2^n) */ | |
b28a02de | 385 | unsigned int gfporder; |
1da177e4 LT |
386 | |
387 | /* force GFP flags, e.g. GFP_DMA */ | |
b28a02de | 388 | gfp_t gfpflags; |
1da177e4 | 389 | |
b28a02de PE |
390 | size_t colour; /* cache colouring range */ |
391 | unsigned int colour_off; /* colour offset */ | |
392 | unsigned int colour_next; /* cache colouring */ | |
393 | kmem_cache_t *slabp_cache; | |
394 | unsigned int slab_size; | |
395 | unsigned int dflags; /* dynamic flags */ | |
1da177e4 LT |
396 | |
397 | /* constructor func */ | |
b28a02de | 398 | void (*ctor) (void *, kmem_cache_t *, unsigned long); |
1da177e4 LT |
399 | |
400 | /* de-constructor func */ | |
b28a02de | 401 | void (*dtor) (void *, kmem_cache_t *, unsigned long); |
1da177e4 LT |
402 | |
403 | /* 4) cache creation/removal */ | |
b28a02de PE |
404 | const char *name; |
405 | struct list_head next; | |
1da177e4 LT |
406 | |
407 | /* 5) statistics */ | |
408 | #if STATS | |
b28a02de PE |
409 | unsigned long num_active; |
410 | unsigned long num_allocations; | |
411 | unsigned long high_mark; | |
412 | unsigned long grown; | |
413 | unsigned long reaped; | |
414 | unsigned long errors; | |
415 | unsigned long max_freeable; | |
416 | unsigned long node_allocs; | |
417 | unsigned long node_frees; | |
418 | atomic_t allochit; | |
419 | atomic_t allocmiss; | |
420 | atomic_t freehit; | |
421 | atomic_t freemiss; | |
1da177e4 LT |
422 | #endif |
423 | #if DEBUG | |
b28a02de PE |
424 | int dbghead; |
425 | int reallen; | |
1da177e4 LT |
426 | #endif |
427 | }; | |
428 | ||
429 | #define CFLGS_OFF_SLAB (0x80000000UL) | |
430 | #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB) | |
431 | ||
432 | #define BATCHREFILL_LIMIT 16 | |
433 | /* Optimization question: fewer reaps means less | |
434 | * probability for unnessary cpucache drain/refill cycles. | |
435 | * | |
dc6f3f27 | 436 | * OTOH the cpuarrays can contain lots of objects, |
1da177e4 LT |
437 | * which could lock up otherwise freeable slabs. |
438 | */ | |
439 | #define REAPTIMEOUT_CPUC (2*HZ) | |
440 | #define REAPTIMEOUT_LIST3 (4*HZ) | |
441 | ||
442 | #if STATS | |
443 | #define STATS_INC_ACTIVE(x) ((x)->num_active++) | |
444 | #define STATS_DEC_ACTIVE(x) ((x)->num_active--) | |
445 | #define STATS_INC_ALLOCED(x) ((x)->num_allocations++) | |
446 | #define STATS_INC_GROWN(x) ((x)->grown++) | |
447 | #define STATS_INC_REAPED(x) ((x)->reaped++) | |
448 | #define STATS_SET_HIGH(x) do { if ((x)->num_active > (x)->high_mark) \ | |
449 | (x)->high_mark = (x)->num_active; \ | |
450 | } while (0) | |
451 | #define STATS_INC_ERR(x) ((x)->errors++) | |
452 | #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++) | |
e498be7d | 453 | #define STATS_INC_NODEFREES(x) ((x)->node_frees++) |
1da177e4 LT |
454 | #define STATS_SET_FREEABLE(x, i) \ |
455 | do { if ((x)->max_freeable < i) \ | |
456 | (x)->max_freeable = i; \ | |
457 | } while (0) | |
458 | ||
459 | #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit) | |
460 | #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss) | |
461 | #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit) | |
462 | #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss) | |
463 | #else | |
464 | #define STATS_INC_ACTIVE(x) do { } while (0) | |
465 | #define STATS_DEC_ACTIVE(x) do { } while (0) | |
466 | #define STATS_INC_ALLOCED(x) do { } while (0) | |
467 | #define STATS_INC_GROWN(x) do { } while (0) | |
468 | #define STATS_INC_REAPED(x) do { } while (0) | |
469 | #define STATS_SET_HIGH(x) do { } while (0) | |
470 | #define STATS_INC_ERR(x) do { } while (0) | |
471 | #define STATS_INC_NODEALLOCS(x) do { } while (0) | |
e498be7d | 472 | #define STATS_INC_NODEFREES(x) do { } while (0) |
1da177e4 LT |
473 | #define STATS_SET_FREEABLE(x, i) \ |
474 | do { } while (0) | |
475 | ||
476 | #define STATS_INC_ALLOCHIT(x) do { } while (0) | |
477 | #define STATS_INC_ALLOCMISS(x) do { } while (0) | |
478 | #define STATS_INC_FREEHIT(x) do { } while (0) | |
479 | #define STATS_INC_FREEMISS(x) do { } while (0) | |
480 | #endif | |
481 | ||
482 | #if DEBUG | |
483 | /* Magic nums for obj red zoning. | |
484 | * Placed in the first word before and the first word after an obj. | |
485 | */ | |
486 | #define RED_INACTIVE 0x5A2CF071UL /* when obj is inactive */ | |
487 | #define RED_ACTIVE 0x170FC2A5UL /* when obj is active */ | |
488 | ||
489 | /* ...and for poisoning */ | |
490 | #define POISON_INUSE 0x5a /* for use-uninitialised poisoning */ | |
491 | #define POISON_FREE 0x6b /* for use-after-free poisoning */ | |
492 | #define POISON_END 0xa5 /* end-byte of poisoning */ | |
493 | ||
494 | /* memory layout of objects: | |
495 | * 0 : objp | |
496 | * 0 .. cachep->dbghead - BYTES_PER_WORD - 1: padding. This ensures that | |
497 | * the end of an object is aligned with the end of the real | |
498 | * allocation. Catches writes behind the end of the allocation. | |
499 | * cachep->dbghead - BYTES_PER_WORD .. cachep->dbghead - 1: | |
500 | * redzone word. | |
501 | * cachep->dbghead: The real object. | |
502 | * cachep->objsize - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long] | |
503 | * cachep->objsize - 1* BYTES_PER_WORD: last caller address [BYTES_PER_WORD long] | |
504 | */ | |
505 | static int obj_dbghead(kmem_cache_t *cachep) | |
506 | { | |
507 | return cachep->dbghead; | |
508 | } | |
509 | ||
510 | static int obj_reallen(kmem_cache_t *cachep) | |
511 | { | |
512 | return cachep->reallen; | |
513 | } | |
514 | ||
515 | static unsigned long *dbg_redzone1(kmem_cache_t *cachep, void *objp) | |
516 | { | |
517 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); | |
518 | return (unsigned long*) (objp+obj_dbghead(cachep)-BYTES_PER_WORD); | |
519 | } | |
520 | ||
521 | static unsigned long *dbg_redzone2(kmem_cache_t *cachep, void *objp) | |
522 | { | |
523 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); | |
524 | if (cachep->flags & SLAB_STORE_USER) | |
b28a02de PE |
525 | return (unsigned long *)(objp + cachep->objsize - |
526 | 2 * BYTES_PER_WORD); | |
527 | return (unsigned long *)(objp + cachep->objsize - BYTES_PER_WORD); | |
1da177e4 LT |
528 | } |
529 | ||
530 | static void **dbg_userword(kmem_cache_t *cachep, void *objp) | |
531 | { | |
532 | BUG_ON(!(cachep->flags & SLAB_STORE_USER)); | |
b28a02de | 533 | return (void **)(objp + cachep->objsize - BYTES_PER_WORD); |
1da177e4 LT |
534 | } |
535 | ||
536 | #else | |
537 | ||
538 | #define obj_dbghead(x) 0 | |
539 | #define obj_reallen(cachep) (cachep->objsize) | |
540 | #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;}) | |
541 | #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;}) | |
542 | #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;}) | |
543 | ||
544 | #endif | |
545 | ||
546 | /* | |
547 | * Maximum size of an obj (in 2^order pages) | |
548 | * and absolute limit for the gfp order. | |
549 | */ | |
550 | #if defined(CONFIG_LARGE_ALLOCS) | |
551 | #define MAX_OBJ_ORDER 13 /* up to 32Mb */ | |
552 | #define MAX_GFP_ORDER 13 /* up to 32Mb */ | |
553 | #elif defined(CONFIG_MMU) | |
554 | #define MAX_OBJ_ORDER 5 /* 32 pages */ | |
555 | #define MAX_GFP_ORDER 5 /* 32 pages */ | |
556 | #else | |
557 | #define MAX_OBJ_ORDER 8 /* up to 1Mb */ | |
558 | #define MAX_GFP_ORDER 8 /* up to 1Mb */ | |
559 | #endif | |
560 | ||
561 | /* | |
562 | * Do not go above this order unless 0 objects fit into the slab. | |
563 | */ | |
564 | #define BREAK_GFP_ORDER_HI 1 | |
565 | #define BREAK_GFP_ORDER_LO 0 | |
566 | static int slab_break_gfp_order = BREAK_GFP_ORDER_LO; | |
567 | ||
065d41cb | 568 | /* Functions for storing/retrieving the cachep and or slab from the |
1da177e4 LT |
569 | * global 'mem_map'. These are used to find the slab an obj belongs to. |
570 | * With kfree(), these are used to find the cache which an obj belongs to. | |
571 | */ | |
065d41cb PE |
572 | static inline void page_set_cache(struct page *page, struct kmem_cache *cache) |
573 | { | |
574 | page->lru.next = (struct list_head *)cache; | |
575 | } | |
576 | ||
577 | static inline struct kmem_cache *page_get_cache(struct page *page) | |
578 | { | |
579 | return (struct kmem_cache *)page->lru.next; | |
580 | } | |
581 | ||
582 | static inline void page_set_slab(struct page *page, struct slab *slab) | |
583 | { | |
584 | page->lru.prev = (struct list_head *)slab; | |
585 | } | |
586 | ||
587 | static inline struct slab *page_get_slab(struct page *page) | |
588 | { | |
589 | return (struct slab *)page->lru.prev; | |
590 | } | |
1da177e4 LT |
591 | |
592 | /* These are the default caches for kmalloc. Custom caches can have other sizes. */ | |
593 | struct cache_sizes malloc_sizes[] = { | |
594 | #define CACHE(x) { .cs_size = (x) }, | |
595 | #include <linux/kmalloc_sizes.h> | |
596 | CACHE(ULONG_MAX) | |
597 | #undef CACHE | |
598 | }; | |
599 | EXPORT_SYMBOL(malloc_sizes); | |
600 | ||
601 | /* Must match cache_sizes above. Out of line to keep cache footprint low. */ | |
602 | struct cache_names { | |
603 | char *name; | |
604 | char *name_dma; | |
605 | }; | |
606 | ||
607 | static struct cache_names __initdata cache_names[] = { | |
608 | #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" }, | |
609 | #include <linux/kmalloc_sizes.h> | |
b28a02de | 610 | {NULL,} |
1da177e4 LT |
611 | #undef CACHE |
612 | }; | |
613 | ||
614 | static struct arraycache_init initarray_cache __initdata = | |
b28a02de | 615 | { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; |
1da177e4 | 616 | static struct arraycache_init initarray_generic = |
b28a02de | 617 | { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; |
1da177e4 LT |
618 | |
619 | /* internal cache of cache description objs */ | |
620 | static kmem_cache_t cache_cache = { | |
b28a02de PE |
621 | .batchcount = 1, |
622 | .limit = BOOT_CPUCACHE_ENTRIES, | |
623 | .shared = 1, | |
624 | .objsize = sizeof(kmem_cache_t), | |
625 | .flags = SLAB_NO_REAP, | |
626 | .spinlock = SPIN_LOCK_UNLOCKED, | |
627 | .name = "kmem_cache", | |
1da177e4 | 628 | #if DEBUG |
b28a02de | 629 | .reallen = sizeof(kmem_cache_t), |
1da177e4 LT |
630 | #endif |
631 | }; | |
632 | ||
633 | /* Guard access to the cache-chain. */ | |
b28a02de | 634 | static struct semaphore cache_chain_sem; |
1da177e4 LT |
635 | static struct list_head cache_chain; |
636 | ||
637 | /* | |
638 | * vm_enough_memory() looks at this to determine how many | |
639 | * slab-allocated pages are possibly freeable under pressure | |
640 | * | |
641 | * SLAB_RECLAIM_ACCOUNT turns this on per-slab | |
642 | */ | |
643 | atomic_t slab_reclaim_pages; | |
1da177e4 LT |
644 | |
645 | /* | |
646 | * chicken and egg problem: delay the per-cpu array allocation | |
647 | * until the general caches are up. | |
648 | */ | |
649 | static enum { | |
650 | NONE, | |
e498be7d CL |
651 | PARTIAL_AC, |
652 | PARTIAL_L3, | |
1da177e4 LT |
653 | FULL |
654 | } g_cpucache_up; | |
655 | ||
656 | static DEFINE_PER_CPU(struct work_struct, reap_work); | |
657 | ||
b28a02de PE |
658 | static void free_block(kmem_cache_t *cachep, void **objpp, int len, int node); |
659 | static void enable_cpucache(kmem_cache_t *cachep); | |
660 | static void cache_reap(void *unused); | |
e498be7d | 661 | static int __node_shrink(kmem_cache_t *cachep, int node); |
1da177e4 LT |
662 | |
663 | static inline struct array_cache *ac_data(kmem_cache_t *cachep) | |
664 | { | |
665 | return cachep->array[smp_processor_id()]; | |
666 | } | |
667 | ||
dd0fc66f | 668 | static inline kmem_cache_t *__find_general_cachep(size_t size, gfp_t gfpflags) |
1da177e4 LT |
669 | { |
670 | struct cache_sizes *csizep = malloc_sizes; | |
671 | ||
672 | #if DEBUG | |
673 | /* This happens if someone tries to call | |
b28a02de PE |
674 | * kmem_cache_create(), or __kmalloc(), before |
675 | * the generic caches are initialized. | |
676 | */ | |
c7e43c78 | 677 | BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL); |
1da177e4 LT |
678 | #endif |
679 | while (size > csizep->cs_size) | |
680 | csizep++; | |
681 | ||
682 | /* | |
0abf40c1 | 683 | * Really subtle: The last entry with cs->cs_size==ULONG_MAX |
1da177e4 LT |
684 | * has cs_{dma,}cachep==NULL. Thus no special case |
685 | * for large kmalloc calls required. | |
686 | */ | |
687 | if (unlikely(gfpflags & GFP_DMA)) | |
688 | return csizep->cs_dmacachep; | |
689 | return csizep->cs_cachep; | |
690 | } | |
691 | ||
dd0fc66f | 692 | kmem_cache_t *kmem_find_general_cachep(size_t size, gfp_t gfpflags) |
97e2bde4 MS |
693 | { |
694 | return __find_general_cachep(size, gfpflags); | |
695 | } | |
696 | EXPORT_SYMBOL(kmem_find_general_cachep); | |
697 | ||
1da177e4 LT |
698 | /* Cal the num objs, wastage, and bytes left over for a given slab size. */ |
699 | static void cache_estimate(unsigned long gfporder, size_t size, size_t align, | |
b28a02de | 700 | int flags, size_t *left_over, unsigned int *num) |
1da177e4 LT |
701 | { |
702 | int i; | |
b28a02de | 703 | size_t wastage = PAGE_SIZE << gfporder; |
1da177e4 LT |
704 | size_t extra = 0; |
705 | size_t base = 0; | |
706 | ||
707 | if (!(flags & CFLGS_OFF_SLAB)) { | |
708 | base = sizeof(struct slab); | |
709 | extra = sizeof(kmem_bufctl_t); | |
710 | } | |
711 | i = 0; | |
b28a02de | 712 | while (i * size + ALIGN(base + i * extra, align) <= wastage) |
1da177e4 LT |
713 | i++; |
714 | if (i > 0) | |
715 | i--; | |
716 | ||
717 | if (i > SLAB_LIMIT) | |
718 | i = SLAB_LIMIT; | |
719 | ||
720 | *num = i; | |
b28a02de PE |
721 | wastage -= i * size; |
722 | wastage -= ALIGN(base + i * extra, align); | |
1da177e4 LT |
723 | *left_over = wastage; |
724 | } | |
725 | ||
726 | #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg) | |
727 | ||
728 | static void __slab_error(const char *function, kmem_cache_t *cachep, char *msg) | |
729 | { | |
730 | printk(KERN_ERR "slab error in %s(): cache `%s': %s\n", | |
b28a02de | 731 | function, cachep->name, msg); |
1da177e4 LT |
732 | dump_stack(); |
733 | } | |
734 | ||
735 | /* | |
736 | * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz | |
737 | * via the workqueue/eventd. | |
738 | * Add the CPU number into the expiration time to minimize the possibility of | |
739 | * the CPUs getting into lockstep and contending for the global cache chain | |
740 | * lock. | |
741 | */ | |
742 | static void __devinit start_cpu_timer(int cpu) | |
743 | { | |
744 | struct work_struct *reap_work = &per_cpu(reap_work, cpu); | |
745 | ||
746 | /* | |
747 | * When this gets called from do_initcalls via cpucache_init(), | |
748 | * init_workqueues() has already run, so keventd will be setup | |
749 | * at that time. | |
750 | */ | |
751 | if (keventd_up() && reap_work->func == NULL) { | |
752 | INIT_WORK(reap_work, cache_reap, NULL); | |
753 | schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu); | |
754 | } | |
755 | } | |
756 | ||
e498be7d | 757 | static struct array_cache *alloc_arraycache(int node, int entries, |
b28a02de | 758 | int batchcount) |
1da177e4 | 759 | { |
b28a02de | 760 | int memsize = sizeof(void *) * entries + sizeof(struct array_cache); |
1da177e4 LT |
761 | struct array_cache *nc = NULL; |
762 | ||
e498be7d | 763 | nc = kmalloc_node(memsize, GFP_KERNEL, node); |
1da177e4 LT |
764 | if (nc) { |
765 | nc->avail = 0; | |
766 | nc->limit = entries; | |
767 | nc->batchcount = batchcount; | |
768 | nc->touched = 0; | |
e498be7d | 769 | spin_lock_init(&nc->lock); |
1da177e4 LT |
770 | } |
771 | return nc; | |
772 | } | |
773 | ||
e498be7d CL |
774 | #ifdef CONFIG_NUMA |
775 | static inline struct array_cache **alloc_alien_cache(int node, int limit) | |
776 | { | |
777 | struct array_cache **ac_ptr; | |
b28a02de | 778 | int memsize = sizeof(void *) * MAX_NUMNODES; |
e498be7d CL |
779 | int i; |
780 | ||
781 | if (limit > 1) | |
782 | limit = 12; | |
783 | ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node); | |
784 | if (ac_ptr) { | |
785 | for_each_node(i) { | |
786 | if (i == node || !node_online(i)) { | |
787 | ac_ptr[i] = NULL; | |
788 | continue; | |
789 | } | |
790 | ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d); | |
791 | if (!ac_ptr[i]) { | |
b28a02de | 792 | for (i--; i <= 0; i--) |
e498be7d CL |
793 | kfree(ac_ptr[i]); |
794 | kfree(ac_ptr); | |
795 | return NULL; | |
796 | } | |
797 | } | |
798 | } | |
799 | return ac_ptr; | |
800 | } | |
801 | ||
802 | static inline void free_alien_cache(struct array_cache **ac_ptr) | |
803 | { | |
804 | int i; | |
805 | ||
806 | if (!ac_ptr) | |
807 | return; | |
808 | ||
809 | for_each_node(i) | |
b28a02de | 810 | kfree(ac_ptr[i]); |
e498be7d CL |
811 | |
812 | kfree(ac_ptr); | |
813 | } | |
814 | ||
b28a02de PE |
815 | static inline void __drain_alien_cache(kmem_cache_t *cachep, |
816 | struct array_cache *ac, int node) | |
e498be7d CL |
817 | { |
818 | struct kmem_list3 *rl3 = cachep->nodelists[node]; | |
819 | ||
820 | if (ac->avail) { | |
821 | spin_lock(&rl3->list_lock); | |
ff69416e | 822 | free_block(cachep, ac->entry, ac->avail, node); |
e498be7d CL |
823 | ac->avail = 0; |
824 | spin_unlock(&rl3->list_lock); | |
825 | } | |
826 | } | |
827 | ||
828 | static void drain_alien_cache(kmem_cache_t *cachep, struct kmem_list3 *l3) | |
829 | { | |
b28a02de | 830 | int i = 0; |
e498be7d CL |
831 | struct array_cache *ac; |
832 | unsigned long flags; | |
833 | ||
834 | for_each_online_node(i) { | |
835 | ac = l3->alien[i]; | |
836 | if (ac) { | |
837 | spin_lock_irqsave(&ac->lock, flags); | |
838 | __drain_alien_cache(cachep, ac, i); | |
839 | spin_unlock_irqrestore(&ac->lock, flags); | |
840 | } | |
841 | } | |
842 | } | |
843 | #else | |
844 | #define alloc_alien_cache(node, limit) do { } while (0) | |
845 | #define free_alien_cache(ac_ptr) do { } while (0) | |
846 | #define drain_alien_cache(cachep, l3) do { } while (0) | |
847 | #endif | |
848 | ||
1da177e4 | 849 | static int __devinit cpuup_callback(struct notifier_block *nfb, |
b28a02de | 850 | unsigned long action, void *hcpu) |
1da177e4 LT |
851 | { |
852 | long cpu = (long)hcpu; | |
b28a02de | 853 | kmem_cache_t *cachep; |
e498be7d CL |
854 | struct kmem_list3 *l3 = NULL; |
855 | int node = cpu_to_node(cpu); | |
856 | int memsize = sizeof(struct kmem_list3); | |
1da177e4 LT |
857 | |
858 | switch (action) { | |
859 | case CPU_UP_PREPARE: | |
860 | down(&cache_chain_sem); | |
e498be7d CL |
861 | /* we need to do this right in the beginning since |
862 | * alloc_arraycache's are going to use this list. | |
863 | * kmalloc_node allows us to add the slab to the right | |
864 | * kmem_list3 and not this cpu's kmem_list3 | |
865 | */ | |
866 | ||
1da177e4 | 867 | list_for_each_entry(cachep, &cache_chain, next) { |
e498be7d CL |
868 | /* setup the size64 kmemlist for cpu before we can |
869 | * begin anything. Make sure some other cpu on this | |
870 | * node has not already allocated this | |
871 | */ | |
872 | if (!cachep->nodelists[node]) { | |
873 | if (!(l3 = kmalloc_node(memsize, | |
b28a02de | 874 | GFP_KERNEL, node))) |
e498be7d CL |
875 | goto bad; |
876 | kmem_list3_init(l3); | |
877 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
b28a02de | 878 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; |
e498be7d CL |
879 | |
880 | cachep->nodelists[node] = l3; | |
881 | } | |
1da177e4 | 882 | |
e498be7d CL |
883 | spin_lock_irq(&cachep->nodelists[node]->list_lock); |
884 | cachep->nodelists[node]->free_limit = | |
b28a02de PE |
885 | (1 + nr_cpus_node(node)) * |
886 | cachep->batchcount + cachep->num; | |
e498be7d CL |
887 | spin_unlock_irq(&cachep->nodelists[node]->list_lock); |
888 | } | |
889 | ||
890 | /* Now we can go ahead with allocating the shared array's | |
b28a02de | 891 | & array cache's */ |
e498be7d | 892 | list_for_each_entry(cachep, &cache_chain, next) { |
cd105df4 TK |
893 | struct array_cache *nc; |
894 | ||
e498be7d | 895 | nc = alloc_arraycache(node, cachep->limit, |
b28a02de | 896 | cachep->batchcount); |
1da177e4 LT |
897 | if (!nc) |
898 | goto bad; | |
1da177e4 | 899 | cachep->array[cpu] = nc; |
1da177e4 | 900 | |
e498be7d CL |
901 | l3 = cachep->nodelists[node]; |
902 | BUG_ON(!l3); | |
903 | if (!l3->shared) { | |
904 | if (!(nc = alloc_arraycache(node, | |
b28a02de PE |
905 | cachep->shared * |
906 | cachep->batchcount, | |
907 | 0xbaadf00d))) | |
908 | goto bad; | |
e498be7d CL |
909 | |
910 | /* we are serialised from CPU_DEAD or | |
b28a02de | 911 | CPU_UP_CANCELLED by the cpucontrol lock */ |
e498be7d CL |
912 | l3->shared = nc; |
913 | } | |
1da177e4 LT |
914 | } |
915 | up(&cache_chain_sem); | |
916 | break; | |
917 | case CPU_ONLINE: | |
918 | start_cpu_timer(cpu); | |
919 | break; | |
920 | #ifdef CONFIG_HOTPLUG_CPU | |
921 | case CPU_DEAD: | |
922 | /* fall thru */ | |
923 | case CPU_UP_CANCELED: | |
924 | down(&cache_chain_sem); | |
925 | ||
926 | list_for_each_entry(cachep, &cache_chain, next) { | |
927 | struct array_cache *nc; | |
e498be7d | 928 | cpumask_t mask; |
1da177e4 | 929 | |
e498be7d | 930 | mask = node_to_cpumask(node); |
1da177e4 LT |
931 | spin_lock_irq(&cachep->spinlock); |
932 | /* cpu is dead; no one can alloc from it. */ | |
933 | nc = cachep->array[cpu]; | |
934 | cachep->array[cpu] = NULL; | |
e498be7d CL |
935 | l3 = cachep->nodelists[node]; |
936 | ||
937 | if (!l3) | |
938 | goto unlock_cache; | |
939 | ||
940 | spin_lock(&l3->list_lock); | |
941 | ||
942 | /* Free limit for this kmem_list3 */ | |
943 | l3->free_limit -= cachep->batchcount; | |
944 | if (nc) | |
ff69416e | 945 | free_block(cachep, nc->entry, nc->avail, node); |
e498be7d CL |
946 | |
947 | if (!cpus_empty(mask)) { | |
b28a02de PE |
948 | spin_unlock(&l3->list_lock); |
949 | goto unlock_cache; | |
950 | } | |
e498be7d CL |
951 | |
952 | if (l3->shared) { | |
953 | free_block(cachep, l3->shared->entry, | |
b28a02de | 954 | l3->shared->avail, node); |
e498be7d CL |
955 | kfree(l3->shared); |
956 | l3->shared = NULL; | |
957 | } | |
958 | if (l3->alien) { | |
959 | drain_alien_cache(cachep, l3); | |
960 | free_alien_cache(l3->alien); | |
961 | l3->alien = NULL; | |
962 | } | |
963 | ||
964 | /* free slabs belonging to this node */ | |
965 | if (__node_shrink(cachep, node)) { | |
966 | cachep->nodelists[node] = NULL; | |
967 | spin_unlock(&l3->list_lock); | |
968 | kfree(l3); | |
969 | } else { | |
970 | spin_unlock(&l3->list_lock); | |
971 | } | |
b28a02de | 972 | unlock_cache: |
1da177e4 LT |
973 | spin_unlock_irq(&cachep->spinlock); |
974 | kfree(nc); | |
975 | } | |
976 | up(&cache_chain_sem); | |
977 | break; | |
978 | #endif | |
979 | } | |
980 | return NOTIFY_OK; | |
b28a02de | 981 | bad: |
1da177e4 LT |
982 | up(&cache_chain_sem); |
983 | return NOTIFY_BAD; | |
984 | } | |
985 | ||
986 | static struct notifier_block cpucache_notifier = { &cpuup_callback, NULL, 0 }; | |
987 | ||
e498be7d CL |
988 | /* |
989 | * swap the static kmem_list3 with kmalloced memory | |
990 | */ | |
b28a02de | 991 | static void init_list(kmem_cache_t *cachep, struct kmem_list3 *list, int nodeid) |
e498be7d CL |
992 | { |
993 | struct kmem_list3 *ptr; | |
994 | ||
995 | BUG_ON(cachep->nodelists[nodeid] != list); | |
996 | ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid); | |
997 | BUG_ON(!ptr); | |
998 | ||
999 | local_irq_disable(); | |
1000 | memcpy(ptr, list, sizeof(struct kmem_list3)); | |
1001 | MAKE_ALL_LISTS(cachep, ptr, nodeid); | |
1002 | cachep->nodelists[nodeid] = ptr; | |
1003 | local_irq_enable(); | |
1004 | } | |
1005 | ||
1da177e4 LT |
1006 | /* Initialisation. |
1007 | * Called after the gfp() functions have been enabled, and before smp_init(). | |
1008 | */ | |
1009 | void __init kmem_cache_init(void) | |
1010 | { | |
1011 | size_t left_over; | |
1012 | struct cache_sizes *sizes; | |
1013 | struct cache_names *names; | |
e498be7d CL |
1014 | int i; |
1015 | ||
1016 | for (i = 0; i < NUM_INIT_LISTS; i++) { | |
1017 | kmem_list3_init(&initkmem_list3[i]); | |
1018 | if (i < MAX_NUMNODES) | |
1019 | cache_cache.nodelists[i] = NULL; | |
1020 | } | |
1da177e4 LT |
1021 | |
1022 | /* | |
1023 | * Fragmentation resistance on low memory - only use bigger | |
1024 | * page orders on machines with more than 32MB of memory. | |
1025 | */ | |
1026 | if (num_physpages > (32 << 20) >> PAGE_SHIFT) | |
1027 | slab_break_gfp_order = BREAK_GFP_ORDER_HI; | |
1028 | ||
1da177e4 LT |
1029 | /* Bootstrap is tricky, because several objects are allocated |
1030 | * from caches that do not exist yet: | |
1031 | * 1) initialize the cache_cache cache: it contains the kmem_cache_t | |
1032 | * structures of all caches, except cache_cache itself: cache_cache | |
1033 | * is statically allocated. | |
e498be7d CL |
1034 | * Initially an __init data area is used for the head array and the |
1035 | * kmem_list3 structures, it's replaced with a kmalloc allocated | |
1036 | * array at the end of the bootstrap. | |
1da177e4 | 1037 | * 2) Create the first kmalloc cache. |
e498be7d CL |
1038 | * The kmem_cache_t for the new cache is allocated normally. |
1039 | * An __init data area is used for the head array. | |
1040 | * 3) Create the remaining kmalloc caches, with minimally sized | |
1041 | * head arrays. | |
1da177e4 LT |
1042 | * 4) Replace the __init data head arrays for cache_cache and the first |
1043 | * kmalloc cache with kmalloc allocated arrays. | |
e498be7d CL |
1044 | * 5) Replace the __init data for kmem_list3 for cache_cache and |
1045 | * the other cache's with kmalloc allocated memory. | |
1046 | * 6) Resize the head arrays of the kmalloc caches to their final sizes. | |
1da177e4 LT |
1047 | */ |
1048 | ||
1049 | /* 1) create the cache_cache */ | |
1050 | init_MUTEX(&cache_chain_sem); | |
1051 | INIT_LIST_HEAD(&cache_chain); | |
1052 | list_add(&cache_cache.next, &cache_chain); | |
1053 | cache_cache.colour_off = cache_line_size(); | |
1054 | cache_cache.array[smp_processor_id()] = &initarray_cache.cache; | |
e498be7d | 1055 | cache_cache.nodelists[numa_node_id()] = &initkmem_list3[CACHE_CACHE]; |
1da177e4 LT |
1056 | |
1057 | cache_cache.objsize = ALIGN(cache_cache.objsize, cache_line_size()); | |
1058 | ||
1059 | cache_estimate(0, cache_cache.objsize, cache_line_size(), 0, | |
b28a02de | 1060 | &left_over, &cache_cache.num); |
1da177e4 LT |
1061 | if (!cache_cache.num) |
1062 | BUG(); | |
1063 | ||
b28a02de | 1064 | cache_cache.colour = left_over / cache_cache.colour_off; |
1da177e4 | 1065 | cache_cache.colour_next = 0; |
b28a02de PE |
1066 | cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) + |
1067 | sizeof(struct slab), cache_line_size()); | |
1da177e4 LT |
1068 | |
1069 | /* 2+3) create the kmalloc caches */ | |
1070 | sizes = malloc_sizes; | |
1071 | names = cache_names; | |
1072 | ||
e498be7d CL |
1073 | /* Initialize the caches that provide memory for the array cache |
1074 | * and the kmem_list3 structures first. | |
1075 | * Without this, further allocations will bug | |
1076 | */ | |
1077 | ||
1078 | sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name, | |
b28a02de PE |
1079 | sizes[INDEX_AC].cs_size, |
1080 | ARCH_KMALLOC_MINALIGN, | |
1081 | (ARCH_KMALLOC_FLAGS | | |
1082 | SLAB_PANIC), NULL, NULL); | |
e498be7d CL |
1083 | |
1084 | if (INDEX_AC != INDEX_L3) | |
1085 | sizes[INDEX_L3].cs_cachep = | |
b28a02de PE |
1086 | kmem_cache_create(names[INDEX_L3].name, |
1087 | sizes[INDEX_L3].cs_size, | |
1088 | ARCH_KMALLOC_MINALIGN, | |
1089 | (ARCH_KMALLOC_FLAGS | SLAB_PANIC), NULL, | |
1090 | NULL); | |
e498be7d | 1091 | |
1da177e4 | 1092 | while (sizes->cs_size != ULONG_MAX) { |
e498be7d CL |
1093 | /* |
1094 | * For performance, all the general caches are L1 aligned. | |
1da177e4 LT |
1095 | * This should be particularly beneficial on SMP boxes, as it |
1096 | * eliminates "false sharing". | |
1097 | * Note for systems short on memory removing the alignment will | |
e498be7d CL |
1098 | * allow tighter packing of the smaller caches. |
1099 | */ | |
b28a02de | 1100 | if (!sizes->cs_cachep) |
e498be7d | 1101 | sizes->cs_cachep = kmem_cache_create(names->name, |
b28a02de PE |
1102 | sizes->cs_size, |
1103 | ARCH_KMALLOC_MINALIGN, | |
1104 | (ARCH_KMALLOC_FLAGS | |
1105 | | SLAB_PANIC), | |
1106 | NULL, NULL); | |
1da177e4 LT |
1107 | |
1108 | /* Inc off-slab bufctl limit until the ceiling is hit. */ | |
1109 | if (!(OFF_SLAB(sizes->cs_cachep))) { | |
b28a02de | 1110 | offslab_limit = sizes->cs_size - sizeof(struct slab); |
1da177e4 LT |
1111 | offslab_limit /= sizeof(kmem_bufctl_t); |
1112 | } | |
1113 | ||
1114 | sizes->cs_dmacachep = kmem_cache_create(names->name_dma, | |
b28a02de PE |
1115 | sizes->cs_size, |
1116 | ARCH_KMALLOC_MINALIGN, | |
1117 | (ARCH_KMALLOC_FLAGS | | |
1118 | SLAB_CACHE_DMA | | |
1119 | SLAB_PANIC), NULL, | |
1120 | NULL); | |
1da177e4 LT |
1121 | |
1122 | sizes++; | |
1123 | names++; | |
1124 | } | |
1125 | /* 4) Replace the bootstrap head arrays */ | |
1126 | { | |
b28a02de | 1127 | void *ptr; |
e498be7d | 1128 | |
1da177e4 | 1129 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); |
e498be7d | 1130 | |
1da177e4 LT |
1131 | local_irq_disable(); |
1132 | BUG_ON(ac_data(&cache_cache) != &initarray_cache.cache); | |
e498be7d | 1133 | memcpy(ptr, ac_data(&cache_cache), |
b28a02de | 1134 | sizeof(struct arraycache_init)); |
1da177e4 LT |
1135 | cache_cache.array[smp_processor_id()] = ptr; |
1136 | local_irq_enable(); | |
e498be7d | 1137 | |
1da177e4 | 1138 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); |
e498be7d | 1139 | |
1da177e4 | 1140 | local_irq_disable(); |
e498be7d | 1141 | BUG_ON(ac_data(malloc_sizes[INDEX_AC].cs_cachep) |
b28a02de | 1142 | != &initarray_generic.cache); |
e498be7d | 1143 | memcpy(ptr, ac_data(malloc_sizes[INDEX_AC].cs_cachep), |
b28a02de | 1144 | sizeof(struct arraycache_init)); |
e498be7d | 1145 | malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] = |
b28a02de | 1146 | ptr; |
1da177e4 LT |
1147 | local_irq_enable(); |
1148 | } | |
e498be7d CL |
1149 | /* 5) Replace the bootstrap kmem_list3's */ |
1150 | { | |
1151 | int node; | |
1152 | /* Replace the static kmem_list3 structures for the boot cpu */ | |
1153 | init_list(&cache_cache, &initkmem_list3[CACHE_CACHE], | |
b28a02de | 1154 | numa_node_id()); |
e498be7d CL |
1155 | |
1156 | for_each_online_node(node) { | |
1157 | init_list(malloc_sizes[INDEX_AC].cs_cachep, | |
b28a02de | 1158 | &initkmem_list3[SIZE_AC + node], node); |
e498be7d CL |
1159 | |
1160 | if (INDEX_AC != INDEX_L3) { | |
1161 | init_list(malloc_sizes[INDEX_L3].cs_cachep, | |
b28a02de PE |
1162 | &initkmem_list3[SIZE_L3 + node], |
1163 | node); | |
e498be7d CL |
1164 | } |
1165 | } | |
1166 | } | |
1da177e4 | 1167 | |
e498be7d | 1168 | /* 6) resize the head arrays to their final sizes */ |
1da177e4 LT |
1169 | { |
1170 | kmem_cache_t *cachep; | |
1171 | down(&cache_chain_sem); | |
1172 | list_for_each_entry(cachep, &cache_chain, next) | |
b28a02de | 1173 | enable_cpucache(cachep); |
1da177e4 LT |
1174 | up(&cache_chain_sem); |
1175 | } | |
1176 | ||
1177 | /* Done! */ | |
1178 | g_cpucache_up = FULL; | |
1179 | ||
1180 | /* Register a cpu startup notifier callback | |
1181 | * that initializes ac_data for all new cpus | |
1182 | */ | |
1183 | register_cpu_notifier(&cpucache_notifier); | |
1da177e4 LT |
1184 | |
1185 | /* The reap timers are started later, with a module init call: | |
1186 | * That part of the kernel is not yet operational. | |
1187 | */ | |
1188 | } | |
1189 | ||
1190 | static int __init cpucache_init(void) | |
1191 | { | |
1192 | int cpu; | |
1193 | ||
1194 | /* | |
1195 | * Register the timers that return unneeded | |
1196 | * pages to gfp. | |
1197 | */ | |
e498be7d | 1198 | for_each_online_cpu(cpu) |
b28a02de | 1199 | start_cpu_timer(cpu); |
1da177e4 LT |
1200 | |
1201 | return 0; | |
1202 | } | |
1203 | ||
1204 | __initcall(cpucache_init); | |
1205 | ||
1206 | /* | |
1207 | * Interface to system's page allocator. No need to hold the cache-lock. | |
1208 | * | |
1209 | * If we requested dmaable memory, we will get it. Even if we | |
1210 | * did not request dmaable memory, we might get it, but that | |
1211 | * would be relatively rare and ignorable. | |
1212 | */ | |
dd0fc66f | 1213 | static void *kmem_getpages(kmem_cache_t *cachep, gfp_t flags, int nodeid) |
1da177e4 LT |
1214 | { |
1215 | struct page *page; | |
1216 | void *addr; | |
1217 | int i; | |
1218 | ||
1219 | flags |= cachep->gfpflags; | |
50c85a19 | 1220 | page = alloc_pages_node(nodeid, flags, cachep->gfporder); |
1da177e4 LT |
1221 | if (!page) |
1222 | return NULL; | |
1223 | addr = page_address(page); | |
1224 | ||
1225 | i = (1 << cachep->gfporder); | |
1226 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) | |
1227 | atomic_add(i, &slab_reclaim_pages); | |
1228 | add_page_state(nr_slab, i); | |
1229 | while (i--) { | |
1230 | SetPageSlab(page); | |
1231 | page++; | |
1232 | } | |
1233 | return addr; | |
1234 | } | |
1235 | ||
1236 | /* | |
1237 | * Interface to system's page release. | |
1238 | */ | |
1239 | static void kmem_freepages(kmem_cache_t *cachep, void *addr) | |
1240 | { | |
b28a02de | 1241 | unsigned long i = (1 << cachep->gfporder); |
1da177e4 LT |
1242 | struct page *page = virt_to_page(addr); |
1243 | const unsigned long nr_freed = i; | |
1244 | ||
1245 | while (i--) { | |
1246 | if (!TestClearPageSlab(page)) | |
1247 | BUG(); | |
1248 | page++; | |
1249 | } | |
1250 | sub_page_state(nr_slab, nr_freed); | |
1251 | if (current->reclaim_state) | |
1252 | current->reclaim_state->reclaimed_slab += nr_freed; | |
1253 | free_pages((unsigned long)addr, cachep->gfporder); | |
b28a02de PE |
1254 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
1255 | atomic_sub(1 << cachep->gfporder, &slab_reclaim_pages); | |
1da177e4 LT |
1256 | } |
1257 | ||
1258 | static void kmem_rcu_free(struct rcu_head *head) | |
1259 | { | |
b28a02de | 1260 | struct slab_rcu *slab_rcu = (struct slab_rcu *)head; |
1da177e4 LT |
1261 | kmem_cache_t *cachep = slab_rcu->cachep; |
1262 | ||
1263 | kmem_freepages(cachep, slab_rcu->addr); | |
1264 | if (OFF_SLAB(cachep)) | |
1265 | kmem_cache_free(cachep->slabp_cache, slab_rcu); | |
1266 | } | |
1267 | ||
1268 | #if DEBUG | |
1269 | ||
1270 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
1271 | static void store_stackinfo(kmem_cache_t *cachep, unsigned long *addr, | |
b28a02de | 1272 | unsigned long caller) |
1da177e4 LT |
1273 | { |
1274 | int size = obj_reallen(cachep); | |
1275 | ||
b28a02de | 1276 | addr = (unsigned long *)&((char *)addr)[obj_dbghead(cachep)]; |
1da177e4 | 1277 | |
b28a02de | 1278 | if (size < 5 * sizeof(unsigned long)) |
1da177e4 LT |
1279 | return; |
1280 | ||
b28a02de PE |
1281 | *addr++ = 0x12345678; |
1282 | *addr++ = caller; | |
1283 | *addr++ = smp_processor_id(); | |
1284 | size -= 3 * sizeof(unsigned long); | |
1da177e4 LT |
1285 | { |
1286 | unsigned long *sptr = &caller; | |
1287 | unsigned long svalue; | |
1288 | ||
1289 | while (!kstack_end(sptr)) { | |
1290 | svalue = *sptr++; | |
1291 | if (kernel_text_address(svalue)) { | |
b28a02de | 1292 | *addr++ = svalue; |
1da177e4 LT |
1293 | size -= sizeof(unsigned long); |
1294 | if (size <= sizeof(unsigned long)) | |
1295 | break; | |
1296 | } | |
1297 | } | |
1298 | ||
1299 | } | |
b28a02de | 1300 | *addr++ = 0x87654321; |
1da177e4 LT |
1301 | } |
1302 | #endif | |
1303 | ||
1304 | static void poison_obj(kmem_cache_t *cachep, void *addr, unsigned char val) | |
1305 | { | |
1306 | int size = obj_reallen(cachep); | |
b28a02de | 1307 | addr = &((char *)addr)[obj_dbghead(cachep)]; |
1da177e4 LT |
1308 | |
1309 | memset(addr, val, size); | |
b28a02de | 1310 | *(unsigned char *)(addr + size - 1) = POISON_END; |
1da177e4 LT |
1311 | } |
1312 | ||
1313 | static void dump_line(char *data, int offset, int limit) | |
1314 | { | |
1315 | int i; | |
1316 | printk(KERN_ERR "%03x:", offset); | |
b28a02de PE |
1317 | for (i = 0; i < limit; i++) { |
1318 | printk(" %02x", (unsigned char)data[offset + i]); | |
1da177e4 LT |
1319 | } |
1320 | printk("\n"); | |
1321 | } | |
1322 | #endif | |
1323 | ||
1324 | #if DEBUG | |
1325 | ||
1326 | static void print_objinfo(kmem_cache_t *cachep, void *objp, int lines) | |
1327 | { | |
1328 | int i, size; | |
1329 | char *realobj; | |
1330 | ||
1331 | if (cachep->flags & SLAB_RED_ZONE) { | |
1332 | printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n", | |
b28a02de PE |
1333 | *dbg_redzone1(cachep, objp), |
1334 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
1335 | } |
1336 | ||
1337 | if (cachep->flags & SLAB_STORE_USER) { | |
1338 | printk(KERN_ERR "Last user: [<%p>]", | |
b28a02de | 1339 | *dbg_userword(cachep, objp)); |
1da177e4 | 1340 | print_symbol("(%s)", |
b28a02de | 1341 | (unsigned long)*dbg_userword(cachep, objp)); |
1da177e4 LT |
1342 | printk("\n"); |
1343 | } | |
b28a02de | 1344 | realobj = (char *)objp + obj_dbghead(cachep); |
1da177e4 | 1345 | size = obj_reallen(cachep); |
b28a02de | 1346 | for (i = 0; i < size && lines; i += 16, lines--) { |
1da177e4 LT |
1347 | int limit; |
1348 | limit = 16; | |
b28a02de PE |
1349 | if (i + limit > size) |
1350 | limit = size - i; | |
1da177e4 LT |
1351 | dump_line(realobj, i, limit); |
1352 | } | |
1353 | } | |
1354 | ||
1355 | static void check_poison_obj(kmem_cache_t *cachep, void *objp) | |
1356 | { | |
1357 | char *realobj; | |
1358 | int size, i; | |
1359 | int lines = 0; | |
1360 | ||
b28a02de | 1361 | realobj = (char *)objp + obj_dbghead(cachep); |
1da177e4 LT |
1362 | size = obj_reallen(cachep); |
1363 | ||
b28a02de | 1364 | for (i = 0; i < size; i++) { |
1da177e4 | 1365 | char exp = POISON_FREE; |
b28a02de | 1366 | if (i == size - 1) |
1da177e4 LT |
1367 | exp = POISON_END; |
1368 | if (realobj[i] != exp) { | |
1369 | int limit; | |
1370 | /* Mismatch ! */ | |
1371 | /* Print header */ | |
1372 | if (lines == 0) { | |
b28a02de PE |
1373 | printk(KERN_ERR |
1374 | "Slab corruption: start=%p, len=%d\n", | |
1375 | realobj, size); | |
1da177e4 LT |
1376 | print_objinfo(cachep, objp, 0); |
1377 | } | |
1378 | /* Hexdump the affected line */ | |
b28a02de | 1379 | i = (i / 16) * 16; |
1da177e4 | 1380 | limit = 16; |
b28a02de PE |
1381 | if (i + limit > size) |
1382 | limit = size - i; | |
1da177e4 LT |
1383 | dump_line(realobj, i, limit); |
1384 | i += 16; | |
1385 | lines++; | |
1386 | /* Limit to 5 lines */ | |
1387 | if (lines > 5) | |
1388 | break; | |
1389 | } | |
1390 | } | |
1391 | if (lines != 0) { | |
1392 | /* Print some data about the neighboring objects, if they | |
1393 | * exist: | |
1394 | */ | |
065d41cb | 1395 | struct slab *slabp = page_get_slab(virt_to_page(objp)); |
1da177e4 LT |
1396 | int objnr; |
1397 | ||
b28a02de | 1398 | objnr = (objp - slabp->s_mem) / cachep->objsize; |
1da177e4 | 1399 | if (objnr) { |
b28a02de PE |
1400 | objp = slabp->s_mem + (objnr - 1) * cachep->objsize; |
1401 | realobj = (char *)objp + obj_dbghead(cachep); | |
1da177e4 | 1402 | printk(KERN_ERR "Prev obj: start=%p, len=%d\n", |
b28a02de | 1403 | realobj, size); |
1da177e4 LT |
1404 | print_objinfo(cachep, objp, 2); |
1405 | } | |
b28a02de PE |
1406 | if (objnr + 1 < cachep->num) { |
1407 | objp = slabp->s_mem + (objnr + 1) * cachep->objsize; | |
1408 | realobj = (char *)objp + obj_dbghead(cachep); | |
1da177e4 | 1409 | printk(KERN_ERR "Next obj: start=%p, len=%d\n", |
b28a02de | 1410 | realobj, size); |
1da177e4 LT |
1411 | print_objinfo(cachep, objp, 2); |
1412 | } | |
1413 | } | |
1414 | } | |
1415 | #endif | |
1416 | ||
1417 | /* Destroy all the objs in a slab, and release the mem back to the system. | |
1418 | * Before calling the slab must have been unlinked from the cache. | |
1419 | * The cache-lock is not held/needed. | |
1420 | */ | |
b28a02de | 1421 | static void slab_destroy(kmem_cache_t *cachep, struct slab *slabp) |
1da177e4 LT |
1422 | { |
1423 | void *addr = slabp->s_mem - slabp->colouroff; | |
1424 | ||
1425 | #if DEBUG | |
1426 | int i; | |
1427 | for (i = 0; i < cachep->num; i++) { | |
1428 | void *objp = slabp->s_mem + cachep->objsize * i; | |
1429 | ||
1430 | if (cachep->flags & SLAB_POISON) { | |
1431 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
b28a02de PE |
1432 | if ((cachep->objsize % PAGE_SIZE) == 0 |
1433 | && OFF_SLAB(cachep)) | |
1434 | kernel_map_pages(virt_to_page(objp), | |
1435 | cachep->objsize / PAGE_SIZE, | |
1436 | 1); | |
1da177e4 LT |
1437 | else |
1438 | check_poison_obj(cachep, objp); | |
1439 | #else | |
1440 | check_poison_obj(cachep, objp); | |
1441 | #endif | |
1442 | } | |
1443 | if (cachep->flags & SLAB_RED_ZONE) { | |
1444 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) | |
1445 | slab_error(cachep, "start of a freed object " | |
b28a02de | 1446 | "was overwritten"); |
1da177e4 LT |
1447 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) |
1448 | slab_error(cachep, "end of a freed object " | |
b28a02de | 1449 | "was overwritten"); |
1da177e4 LT |
1450 | } |
1451 | if (cachep->dtor && !(cachep->flags & SLAB_POISON)) | |
b28a02de | 1452 | (cachep->dtor) (objp + obj_dbghead(cachep), cachep, 0); |
1da177e4 LT |
1453 | } |
1454 | #else | |
1455 | if (cachep->dtor) { | |
1456 | int i; | |
1457 | for (i = 0; i < cachep->num; i++) { | |
b28a02de PE |
1458 | void *objp = slabp->s_mem + cachep->objsize * i; |
1459 | (cachep->dtor) (objp, cachep, 0); | |
1da177e4 LT |
1460 | } |
1461 | } | |
1462 | #endif | |
1463 | ||
1464 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) { | |
1465 | struct slab_rcu *slab_rcu; | |
1466 | ||
b28a02de | 1467 | slab_rcu = (struct slab_rcu *)slabp; |
1da177e4 LT |
1468 | slab_rcu->cachep = cachep; |
1469 | slab_rcu->addr = addr; | |
1470 | call_rcu(&slab_rcu->head, kmem_rcu_free); | |
1471 | } else { | |
1472 | kmem_freepages(cachep, addr); | |
1473 | if (OFF_SLAB(cachep)) | |
1474 | kmem_cache_free(cachep->slabp_cache, slabp); | |
1475 | } | |
1476 | } | |
1477 | ||
e498be7d CL |
1478 | /* For setting up all the kmem_list3s for cache whose objsize is same |
1479 | as size of kmem_list3. */ | |
1480 | static inline void set_up_list3s(kmem_cache_t *cachep, int index) | |
1481 | { | |
1482 | int node; | |
1483 | ||
1484 | for_each_online_node(node) { | |
b28a02de | 1485 | cachep->nodelists[node] = &initkmem_list3[index + node]; |
e498be7d | 1486 | cachep->nodelists[node]->next_reap = jiffies + |
b28a02de PE |
1487 | REAPTIMEOUT_LIST3 + |
1488 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
e498be7d CL |
1489 | } |
1490 | } | |
1491 | ||
4d268eba PE |
1492 | /** |
1493 | * calculate_slab_order - calculate size (page order) of slabs and the number | |
1494 | * of objects per slab. | |
1495 | * | |
1496 | * This could be made much more intelligent. For now, try to avoid using | |
1497 | * high order pages for slabs. When the gfp() functions are more friendly | |
1498 | * towards high-order requests, this should be changed. | |
1499 | */ | |
1500 | static inline size_t calculate_slab_order(kmem_cache_t *cachep, size_t size, | |
1501 | size_t align, gfp_t flags) | |
1502 | { | |
1503 | size_t left_over = 0; | |
1504 | ||
b28a02de | 1505 | for (;; cachep->gfporder++) { |
4d268eba PE |
1506 | unsigned int num; |
1507 | size_t remainder; | |
1508 | ||
1509 | if (cachep->gfporder > MAX_GFP_ORDER) { | |
1510 | cachep->num = 0; | |
1511 | break; | |
1512 | } | |
1513 | ||
1514 | cache_estimate(cachep->gfporder, size, align, flags, | |
1515 | &remainder, &num); | |
1516 | if (!num) | |
1517 | continue; | |
1518 | /* More than offslab_limit objects will cause problems */ | |
1519 | if (flags & CFLGS_OFF_SLAB && cachep->num > offslab_limit) | |
1520 | break; | |
1521 | ||
1522 | cachep->num = num; | |
1523 | left_over = remainder; | |
1524 | ||
1525 | /* | |
1526 | * Large number of objects is good, but very large slabs are | |
1527 | * currently bad for the gfp()s. | |
1528 | */ | |
1529 | if (cachep->gfporder >= slab_break_gfp_order) | |
1530 | break; | |
1531 | ||
1532 | if ((left_over * 8) <= (PAGE_SIZE << cachep->gfporder)) | |
1533 | /* Acceptable internal fragmentation */ | |
1534 | break; | |
1535 | } | |
1536 | return left_over; | |
1537 | } | |
1538 | ||
1da177e4 LT |
1539 | /** |
1540 | * kmem_cache_create - Create a cache. | |
1541 | * @name: A string which is used in /proc/slabinfo to identify this cache. | |
1542 | * @size: The size of objects to be created in this cache. | |
1543 | * @align: The required alignment for the objects. | |
1544 | * @flags: SLAB flags | |
1545 | * @ctor: A constructor for the objects. | |
1546 | * @dtor: A destructor for the objects. | |
1547 | * | |
1548 | * Returns a ptr to the cache on success, NULL on failure. | |
1549 | * Cannot be called within a int, but can be interrupted. | |
1550 | * The @ctor is run when new pages are allocated by the cache | |
1551 | * and the @dtor is run before the pages are handed back. | |
1552 | * | |
1553 | * @name must be valid until the cache is destroyed. This implies that | |
1554 | * the module calling this has to destroy the cache before getting | |
1555 | * unloaded. | |
1556 | * | |
1557 | * The flags are | |
1558 | * | |
1559 | * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5) | |
1560 | * to catch references to uninitialised memory. | |
1561 | * | |
1562 | * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check | |
1563 | * for buffer overruns. | |
1564 | * | |
1565 | * %SLAB_NO_REAP - Don't automatically reap this cache when we're under | |
1566 | * memory pressure. | |
1567 | * | |
1568 | * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware | |
1569 | * cacheline. This can be beneficial if you're counting cycles as closely | |
1570 | * as davem. | |
1571 | */ | |
1572 | kmem_cache_t * | |
1573 | kmem_cache_create (const char *name, size_t size, size_t align, | |
1574 | unsigned long flags, void (*ctor)(void*, kmem_cache_t *, unsigned long), | |
1575 | void (*dtor)(void*, kmem_cache_t *, unsigned long)) | |
1576 | { | |
1577 | size_t left_over, slab_size, ralign; | |
1578 | kmem_cache_t *cachep = NULL; | |
4f12bb4f | 1579 | struct list_head *p; |
1da177e4 LT |
1580 | |
1581 | /* | |
1582 | * Sanity checks... these are all serious usage bugs. | |
1583 | */ | |
1584 | if ((!name) || | |
b28a02de PE |
1585 | in_interrupt() || |
1586 | (size < BYTES_PER_WORD) || | |
1587 | (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) { | |
1588 | printk(KERN_ERR "%s: Early error in slab %s\n", | |
1589 | __FUNCTION__, name); | |
1590 | BUG(); | |
1591 | } | |
1da177e4 | 1592 | |
4f12bb4f AM |
1593 | down(&cache_chain_sem); |
1594 | ||
1595 | list_for_each(p, &cache_chain) { | |
1596 | kmem_cache_t *pc = list_entry(p, kmem_cache_t, next); | |
1597 | mm_segment_t old_fs = get_fs(); | |
1598 | char tmp; | |
1599 | int res; | |
1600 | ||
1601 | /* | |
1602 | * This happens when the module gets unloaded and doesn't | |
1603 | * destroy its slab cache and no-one else reuses the vmalloc | |
1604 | * area of the module. Print a warning. | |
1605 | */ | |
1606 | set_fs(KERNEL_DS); | |
1607 | res = __get_user(tmp, pc->name); | |
1608 | set_fs(old_fs); | |
1609 | if (res) { | |
1610 | printk("SLAB: cache with size %d has lost its name\n", | |
b28a02de | 1611 | pc->objsize); |
4f12bb4f AM |
1612 | continue; |
1613 | } | |
1614 | ||
b28a02de | 1615 | if (!strcmp(pc->name, name)) { |
4f12bb4f AM |
1616 | printk("kmem_cache_create: duplicate cache %s\n", name); |
1617 | dump_stack(); | |
1618 | goto oops; | |
1619 | } | |
1620 | } | |
1621 | ||
1da177e4 LT |
1622 | #if DEBUG |
1623 | WARN_ON(strchr(name, ' ')); /* It confuses parsers */ | |
1624 | if ((flags & SLAB_DEBUG_INITIAL) && !ctor) { | |
1625 | /* No constructor, but inital state check requested */ | |
1626 | printk(KERN_ERR "%s: No con, but init state check " | |
b28a02de | 1627 | "requested - %s\n", __FUNCTION__, name); |
1da177e4 LT |
1628 | flags &= ~SLAB_DEBUG_INITIAL; |
1629 | } | |
1da177e4 LT |
1630 | #if FORCED_DEBUG |
1631 | /* | |
1632 | * Enable redzoning and last user accounting, except for caches with | |
1633 | * large objects, if the increased size would increase the object size | |
1634 | * above the next power of two: caches with object sizes just above a | |
1635 | * power of two have a significant amount of internal fragmentation. | |
1636 | */ | |
b28a02de PE |
1637 | if ((size < 4096 |
1638 | || fls(size - 1) == fls(size - 1 + 3 * BYTES_PER_WORD))) | |
1639 | flags |= SLAB_RED_ZONE | SLAB_STORE_USER; | |
1da177e4 LT |
1640 | if (!(flags & SLAB_DESTROY_BY_RCU)) |
1641 | flags |= SLAB_POISON; | |
1642 | #endif | |
1643 | if (flags & SLAB_DESTROY_BY_RCU) | |
1644 | BUG_ON(flags & SLAB_POISON); | |
1645 | #endif | |
1646 | if (flags & SLAB_DESTROY_BY_RCU) | |
1647 | BUG_ON(dtor); | |
1648 | ||
1649 | /* | |
1650 | * Always checks flags, a caller might be expecting debug | |
1651 | * support which isn't available. | |
1652 | */ | |
1653 | if (flags & ~CREATE_MASK) | |
1654 | BUG(); | |
1655 | ||
1656 | /* Check that size is in terms of words. This is needed to avoid | |
1657 | * unaligned accesses for some archs when redzoning is used, and makes | |
1658 | * sure any on-slab bufctl's are also correctly aligned. | |
1659 | */ | |
b28a02de PE |
1660 | if (size & (BYTES_PER_WORD - 1)) { |
1661 | size += (BYTES_PER_WORD - 1); | |
1662 | size &= ~(BYTES_PER_WORD - 1); | |
1da177e4 LT |
1663 | } |
1664 | ||
1665 | /* calculate out the final buffer alignment: */ | |
1666 | /* 1) arch recommendation: can be overridden for debug */ | |
1667 | if (flags & SLAB_HWCACHE_ALIGN) { | |
1668 | /* Default alignment: as specified by the arch code. | |
1669 | * Except if an object is really small, then squeeze multiple | |
1670 | * objects into one cacheline. | |
1671 | */ | |
1672 | ralign = cache_line_size(); | |
b28a02de | 1673 | while (size <= ralign / 2) |
1da177e4 LT |
1674 | ralign /= 2; |
1675 | } else { | |
1676 | ralign = BYTES_PER_WORD; | |
1677 | } | |
1678 | /* 2) arch mandated alignment: disables debug if necessary */ | |
1679 | if (ralign < ARCH_SLAB_MINALIGN) { | |
1680 | ralign = ARCH_SLAB_MINALIGN; | |
1681 | if (ralign > BYTES_PER_WORD) | |
b28a02de | 1682 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); |
1da177e4 LT |
1683 | } |
1684 | /* 3) caller mandated alignment: disables debug if necessary */ | |
1685 | if (ralign < align) { | |
1686 | ralign = align; | |
1687 | if (ralign > BYTES_PER_WORD) | |
b28a02de | 1688 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); |
1da177e4 LT |
1689 | } |
1690 | /* 4) Store it. Note that the debug code below can reduce | |
1691 | * the alignment to BYTES_PER_WORD. | |
1692 | */ | |
1693 | align = ralign; | |
1694 | ||
1695 | /* Get cache's description obj. */ | |
1696 | cachep = (kmem_cache_t *) kmem_cache_alloc(&cache_cache, SLAB_KERNEL); | |
1697 | if (!cachep) | |
4f12bb4f | 1698 | goto oops; |
1da177e4 LT |
1699 | memset(cachep, 0, sizeof(kmem_cache_t)); |
1700 | ||
1701 | #if DEBUG | |
1702 | cachep->reallen = size; | |
1703 | ||
1704 | if (flags & SLAB_RED_ZONE) { | |
1705 | /* redzoning only works with word aligned caches */ | |
1706 | align = BYTES_PER_WORD; | |
1707 | ||
1708 | /* add space for red zone words */ | |
1709 | cachep->dbghead += BYTES_PER_WORD; | |
b28a02de | 1710 | size += 2 * BYTES_PER_WORD; |
1da177e4 LT |
1711 | } |
1712 | if (flags & SLAB_STORE_USER) { | |
1713 | /* user store requires word alignment and | |
1714 | * one word storage behind the end of the real | |
1715 | * object. | |
1716 | */ | |
1717 | align = BYTES_PER_WORD; | |
1718 | size += BYTES_PER_WORD; | |
1719 | } | |
1720 | #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC) | |
b28a02de PE |
1721 | if (size >= malloc_sizes[INDEX_L3 + 1].cs_size |
1722 | && cachep->reallen > cache_line_size() && size < PAGE_SIZE) { | |
1da177e4 LT |
1723 | cachep->dbghead += PAGE_SIZE - size; |
1724 | size = PAGE_SIZE; | |
1725 | } | |
1726 | #endif | |
1727 | #endif | |
1728 | ||
1729 | /* Determine if the slab management is 'on' or 'off' slab. */ | |
b28a02de | 1730 | if (size >= (PAGE_SIZE >> 3)) |
1da177e4 LT |
1731 | /* |
1732 | * Size is large, assume best to place the slab management obj | |
1733 | * off-slab (should allow better packing of objs). | |
1734 | */ | |
1735 | flags |= CFLGS_OFF_SLAB; | |
1736 | ||
1737 | size = ALIGN(size, align); | |
1738 | ||
1739 | if ((flags & SLAB_RECLAIM_ACCOUNT) && size <= PAGE_SIZE) { | |
1740 | /* | |
1741 | * A VFS-reclaimable slab tends to have most allocations | |
1742 | * as GFP_NOFS and we really don't want to have to be allocating | |
1743 | * higher-order pages when we are unable to shrink dcache. | |
1744 | */ | |
1745 | cachep->gfporder = 0; | |
1746 | cache_estimate(cachep->gfporder, size, align, flags, | |
b28a02de | 1747 | &left_over, &cachep->num); |
4d268eba PE |
1748 | } else |
1749 | left_over = calculate_slab_order(cachep, size, align, flags); | |
1da177e4 LT |
1750 | |
1751 | if (!cachep->num) { | |
1752 | printk("kmem_cache_create: couldn't create cache %s.\n", name); | |
1753 | kmem_cache_free(&cache_cache, cachep); | |
1754 | cachep = NULL; | |
4f12bb4f | 1755 | goto oops; |
1da177e4 | 1756 | } |
b28a02de PE |
1757 | slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t) |
1758 | + sizeof(struct slab), align); | |
1da177e4 LT |
1759 | |
1760 | /* | |
1761 | * If the slab has been placed off-slab, and we have enough space then | |
1762 | * move it on-slab. This is at the expense of any extra colouring. | |
1763 | */ | |
1764 | if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) { | |
1765 | flags &= ~CFLGS_OFF_SLAB; | |
1766 | left_over -= slab_size; | |
1767 | } | |
1768 | ||
1769 | if (flags & CFLGS_OFF_SLAB) { | |
1770 | /* really off slab. No need for manual alignment */ | |
b28a02de PE |
1771 | slab_size = |
1772 | cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab); | |
1da177e4 LT |
1773 | } |
1774 | ||
1775 | cachep->colour_off = cache_line_size(); | |
1776 | /* Offset must be a multiple of the alignment. */ | |
1777 | if (cachep->colour_off < align) | |
1778 | cachep->colour_off = align; | |
b28a02de | 1779 | cachep->colour = left_over / cachep->colour_off; |
1da177e4 LT |
1780 | cachep->slab_size = slab_size; |
1781 | cachep->flags = flags; | |
1782 | cachep->gfpflags = 0; | |
1783 | if (flags & SLAB_CACHE_DMA) | |
1784 | cachep->gfpflags |= GFP_DMA; | |
1785 | spin_lock_init(&cachep->spinlock); | |
1786 | cachep->objsize = size; | |
1da177e4 LT |
1787 | |
1788 | if (flags & CFLGS_OFF_SLAB) | |
b2d55073 | 1789 | cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u); |
1da177e4 LT |
1790 | cachep->ctor = ctor; |
1791 | cachep->dtor = dtor; | |
1792 | cachep->name = name; | |
1793 | ||
1794 | /* Don't let CPUs to come and go */ | |
1795 | lock_cpu_hotplug(); | |
1796 | ||
1797 | if (g_cpucache_up == FULL) { | |
1798 | enable_cpucache(cachep); | |
1799 | } else { | |
1800 | if (g_cpucache_up == NONE) { | |
1801 | /* Note: the first kmem_cache_create must create | |
1802 | * the cache that's used by kmalloc(24), otherwise | |
1803 | * the creation of further caches will BUG(). | |
1804 | */ | |
e498be7d | 1805 | cachep->array[smp_processor_id()] = |
b28a02de | 1806 | &initarray_generic.cache; |
e498be7d CL |
1807 | |
1808 | /* If the cache that's used by | |
1809 | * kmalloc(sizeof(kmem_list3)) is the first cache, | |
1810 | * then we need to set up all its list3s, otherwise | |
1811 | * the creation of further caches will BUG(). | |
1812 | */ | |
1813 | set_up_list3s(cachep, SIZE_AC); | |
1814 | if (INDEX_AC == INDEX_L3) | |
1815 | g_cpucache_up = PARTIAL_L3; | |
1816 | else | |
1817 | g_cpucache_up = PARTIAL_AC; | |
1da177e4 | 1818 | } else { |
e498be7d | 1819 | cachep->array[smp_processor_id()] = |
b28a02de | 1820 | kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); |
e498be7d CL |
1821 | |
1822 | if (g_cpucache_up == PARTIAL_AC) { | |
1823 | set_up_list3s(cachep, SIZE_L3); | |
1824 | g_cpucache_up = PARTIAL_L3; | |
1825 | } else { | |
1826 | int node; | |
1827 | for_each_online_node(node) { | |
1828 | ||
1829 | cachep->nodelists[node] = | |
b28a02de PE |
1830 | kmalloc_node(sizeof |
1831 | (struct kmem_list3), | |
1832 | GFP_KERNEL, node); | |
e498be7d | 1833 | BUG_ON(!cachep->nodelists[node]); |
b28a02de PE |
1834 | kmem_list3_init(cachep-> |
1835 | nodelists[node]); | |
e498be7d CL |
1836 | } |
1837 | } | |
1da177e4 | 1838 | } |
e498be7d | 1839 | cachep->nodelists[numa_node_id()]->next_reap = |
b28a02de PE |
1840 | jiffies + REAPTIMEOUT_LIST3 + |
1841 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
e498be7d | 1842 | |
1da177e4 LT |
1843 | BUG_ON(!ac_data(cachep)); |
1844 | ac_data(cachep)->avail = 0; | |
1845 | ac_data(cachep)->limit = BOOT_CPUCACHE_ENTRIES; | |
1846 | ac_data(cachep)->batchcount = 1; | |
1847 | ac_data(cachep)->touched = 0; | |
1848 | cachep->batchcount = 1; | |
1849 | cachep->limit = BOOT_CPUCACHE_ENTRIES; | |
b28a02de | 1850 | } |
1da177e4 | 1851 | |
1da177e4 LT |
1852 | /* cache setup completed, link it into the list */ |
1853 | list_add(&cachep->next, &cache_chain); | |
1da177e4 | 1854 | unlock_cpu_hotplug(); |
b28a02de | 1855 | oops: |
1da177e4 LT |
1856 | if (!cachep && (flags & SLAB_PANIC)) |
1857 | panic("kmem_cache_create(): failed to create slab `%s'\n", | |
b28a02de | 1858 | name); |
4f12bb4f | 1859 | up(&cache_chain_sem); |
1da177e4 LT |
1860 | return cachep; |
1861 | } | |
1862 | EXPORT_SYMBOL(kmem_cache_create); | |
1863 | ||
1864 | #if DEBUG | |
1865 | static void check_irq_off(void) | |
1866 | { | |
1867 | BUG_ON(!irqs_disabled()); | |
1868 | } | |
1869 | ||
1870 | static void check_irq_on(void) | |
1871 | { | |
1872 | BUG_ON(irqs_disabled()); | |
1873 | } | |
1874 | ||
1875 | static void check_spinlock_acquired(kmem_cache_t *cachep) | |
1876 | { | |
1877 | #ifdef CONFIG_SMP | |
1878 | check_irq_off(); | |
e498be7d | 1879 | assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock); |
1da177e4 LT |
1880 | #endif |
1881 | } | |
e498be7d CL |
1882 | |
1883 | static inline void check_spinlock_acquired_node(kmem_cache_t *cachep, int node) | |
1884 | { | |
1885 | #ifdef CONFIG_SMP | |
1886 | check_irq_off(); | |
1887 | assert_spin_locked(&cachep->nodelists[node]->list_lock); | |
1888 | #endif | |
1889 | } | |
1890 | ||
1da177e4 LT |
1891 | #else |
1892 | #define check_irq_off() do { } while(0) | |
1893 | #define check_irq_on() do { } while(0) | |
1894 | #define check_spinlock_acquired(x) do { } while(0) | |
e498be7d | 1895 | #define check_spinlock_acquired_node(x, y) do { } while(0) |
1da177e4 LT |
1896 | #endif |
1897 | ||
1898 | /* | |
1899 | * Waits for all CPUs to execute func(). | |
1900 | */ | |
b28a02de | 1901 | static void smp_call_function_all_cpus(void (*func)(void *arg), void *arg) |
1da177e4 LT |
1902 | { |
1903 | check_irq_on(); | |
1904 | preempt_disable(); | |
1905 | ||
1906 | local_irq_disable(); | |
1907 | func(arg); | |
1908 | local_irq_enable(); | |
1909 | ||
1910 | if (smp_call_function(func, arg, 1, 1)) | |
1911 | BUG(); | |
1912 | ||
1913 | preempt_enable(); | |
1914 | } | |
1915 | ||
b28a02de PE |
1916 | static void drain_array_locked(kmem_cache_t *cachep, struct array_cache *ac, |
1917 | int force, int node); | |
1da177e4 LT |
1918 | |
1919 | static void do_drain(void *arg) | |
1920 | { | |
b28a02de | 1921 | kmem_cache_t *cachep = (kmem_cache_t *) arg; |
1da177e4 | 1922 | struct array_cache *ac; |
ff69416e | 1923 | int node = numa_node_id(); |
1da177e4 LT |
1924 | |
1925 | check_irq_off(); | |
1926 | ac = ac_data(cachep); | |
ff69416e CL |
1927 | spin_lock(&cachep->nodelists[node]->list_lock); |
1928 | free_block(cachep, ac->entry, ac->avail, node); | |
1929 | spin_unlock(&cachep->nodelists[node]->list_lock); | |
1da177e4 LT |
1930 | ac->avail = 0; |
1931 | } | |
1932 | ||
1933 | static void drain_cpu_caches(kmem_cache_t *cachep) | |
1934 | { | |
e498be7d CL |
1935 | struct kmem_list3 *l3; |
1936 | int node; | |
1937 | ||
1da177e4 LT |
1938 | smp_call_function_all_cpus(do_drain, cachep); |
1939 | check_irq_on(); | |
1940 | spin_lock_irq(&cachep->spinlock); | |
b28a02de | 1941 | for_each_online_node(node) { |
e498be7d CL |
1942 | l3 = cachep->nodelists[node]; |
1943 | if (l3) { | |
1944 | spin_lock(&l3->list_lock); | |
1945 | drain_array_locked(cachep, l3->shared, 1, node); | |
1946 | spin_unlock(&l3->list_lock); | |
1947 | if (l3->alien) | |
1948 | drain_alien_cache(cachep, l3); | |
1949 | } | |
1950 | } | |
1da177e4 LT |
1951 | spin_unlock_irq(&cachep->spinlock); |
1952 | } | |
1953 | ||
e498be7d | 1954 | static int __node_shrink(kmem_cache_t *cachep, int node) |
1da177e4 LT |
1955 | { |
1956 | struct slab *slabp; | |
e498be7d | 1957 | struct kmem_list3 *l3 = cachep->nodelists[node]; |
1da177e4 LT |
1958 | int ret; |
1959 | ||
e498be7d | 1960 | for (;;) { |
1da177e4 LT |
1961 | struct list_head *p; |
1962 | ||
e498be7d CL |
1963 | p = l3->slabs_free.prev; |
1964 | if (p == &l3->slabs_free) | |
1da177e4 LT |
1965 | break; |
1966 | ||
e498be7d | 1967 | slabp = list_entry(l3->slabs_free.prev, struct slab, list); |
1da177e4 LT |
1968 | #if DEBUG |
1969 | if (slabp->inuse) | |
1970 | BUG(); | |
1971 | #endif | |
1972 | list_del(&slabp->list); | |
1973 | ||
e498be7d CL |
1974 | l3->free_objects -= cachep->num; |
1975 | spin_unlock_irq(&l3->list_lock); | |
1da177e4 | 1976 | slab_destroy(cachep, slabp); |
e498be7d | 1977 | spin_lock_irq(&l3->list_lock); |
1da177e4 | 1978 | } |
b28a02de | 1979 | ret = !list_empty(&l3->slabs_full) || !list_empty(&l3->slabs_partial); |
1da177e4 LT |
1980 | return ret; |
1981 | } | |
1982 | ||
e498be7d CL |
1983 | static int __cache_shrink(kmem_cache_t *cachep) |
1984 | { | |
1985 | int ret = 0, i = 0; | |
1986 | struct kmem_list3 *l3; | |
1987 | ||
1988 | drain_cpu_caches(cachep); | |
1989 | ||
1990 | check_irq_on(); | |
1991 | for_each_online_node(i) { | |
1992 | l3 = cachep->nodelists[i]; | |
1993 | if (l3) { | |
1994 | spin_lock_irq(&l3->list_lock); | |
1995 | ret += __node_shrink(cachep, i); | |
1996 | spin_unlock_irq(&l3->list_lock); | |
1997 | } | |
1998 | } | |
1999 | return (ret ? 1 : 0); | |
2000 | } | |
2001 | ||
1da177e4 LT |
2002 | /** |
2003 | * kmem_cache_shrink - Shrink a cache. | |
2004 | * @cachep: The cache to shrink. | |
2005 | * | |
2006 | * Releases as many slabs as possible for a cache. | |
2007 | * To help debugging, a zero exit status indicates all slabs were released. | |
2008 | */ | |
2009 | int kmem_cache_shrink(kmem_cache_t *cachep) | |
2010 | { | |
2011 | if (!cachep || in_interrupt()) | |
2012 | BUG(); | |
2013 | ||
2014 | return __cache_shrink(cachep); | |
2015 | } | |
2016 | EXPORT_SYMBOL(kmem_cache_shrink); | |
2017 | ||
2018 | /** | |
2019 | * kmem_cache_destroy - delete a cache | |
2020 | * @cachep: the cache to destroy | |
2021 | * | |
2022 | * Remove a kmem_cache_t object from the slab cache. | |
2023 | * Returns 0 on success. | |
2024 | * | |
2025 | * It is expected this function will be called by a module when it is | |
2026 | * unloaded. This will remove the cache completely, and avoid a duplicate | |
2027 | * cache being allocated each time a module is loaded and unloaded, if the | |
2028 | * module doesn't have persistent in-kernel storage across loads and unloads. | |
2029 | * | |
2030 | * The cache must be empty before calling this function. | |
2031 | * | |
2032 | * The caller must guarantee that noone will allocate memory from the cache | |
2033 | * during the kmem_cache_destroy(). | |
2034 | */ | |
b28a02de | 2035 | int kmem_cache_destroy(kmem_cache_t *cachep) |
1da177e4 LT |
2036 | { |
2037 | int i; | |
e498be7d | 2038 | struct kmem_list3 *l3; |
1da177e4 LT |
2039 | |
2040 | if (!cachep || in_interrupt()) | |
2041 | BUG(); | |
2042 | ||
2043 | /* Don't let CPUs to come and go */ | |
2044 | lock_cpu_hotplug(); | |
2045 | ||
2046 | /* Find the cache in the chain of caches. */ | |
2047 | down(&cache_chain_sem); | |
2048 | /* | |
2049 | * the chain is never empty, cache_cache is never destroyed | |
2050 | */ | |
2051 | list_del(&cachep->next); | |
2052 | up(&cache_chain_sem); | |
2053 | ||
2054 | if (__cache_shrink(cachep)) { | |
2055 | slab_error(cachep, "Can't free all objects"); | |
2056 | down(&cache_chain_sem); | |
b28a02de | 2057 | list_add(&cachep->next, &cache_chain); |
1da177e4 LT |
2058 | up(&cache_chain_sem); |
2059 | unlock_cpu_hotplug(); | |
2060 | return 1; | |
2061 | } | |
2062 | ||
2063 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) | |
fbd568a3 | 2064 | synchronize_rcu(); |
1da177e4 | 2065 | |
e498be7d | 2066 | for_each_online_cpu(i) |
b28a02de | 2067 | kfree(cachep->array[i]); |
1da177e4 LT |
2068 | |
2069 | /* NUMA: free the list3 structures */ | |
e498be7d CL |
2070 | for_each_online_node(i) { |
2071 | if ((l3 = cachep->nodelists[i])) { | |
2072 | kfree(l3->shared); | |
2073 | free_alien_cache(l3->alien); | |
2074 | kfree(l3); | |
2075 | } | |
2076 | } | |
1da177e4 LT |
2077 | kmem_cache_free(&cache_cache, cachep); |
2078 | ||
2079 | unlock_cpu_hotplug(); | |
2080 | ||
2081 | return 0; | |
2082 | } | |
2083 | EXPORT_SYMBOL(kmem_cache_destroy); | |
2084 | ||
2085 | /* Get the memory for a slab management obj. */ | |
b28a02de PE |
2086 | static struct slab *alloc_slabmgmt(kmem_cache_t *cachep, void *objp, |
2087 | int colour_off, gfp_t local_flags) | |
1da177e4 LT |
2088 | { |
2089 | struct slab *slabp; | |
b28a02de | 2090 | |
1da177e4 LT |
2091 | if (OFF_SLAB(cachep)) { |
2092 | /* Slab management obj is off-slab. */ | |
2093 | slabp = kmem_cache_alloc(cachep->slabp_cache, local_flags); | |
2094 | if (!slabp) | |
2095 | return NULL; | |
2096 | } else { | |
b28a02de | 2097 | slabp = objp + colour_off; |
1da177e4 LT |
2098 | colour_off += cachep->slab_size; |
2099 | } | |
2100 | slabp->inuse = 0; | |
2101 | slabp->colouroff = colour_off; | |
b28a02de | 2102 | slabp->s_mem = objp + colour_off; |
1da177e4 LT |
2103 | |
2104 | return slabp; | |
2105 | } | |
2106 | ||
2107 | static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp) | |
2108 | { | |
b28a02de | 2109 | return (kmem_bufctl_t *) (slabp + 1); |
1da177e4 LT |
2110 | } |
2111 | ||
2112 | static void cache_init_objs(kmem_cache_t *cachep, | |
b28a02de | 2113 | struct slab *slabp, unsigned long ctor_flags) |
1da177e4 LT |
2114 | { |
2115 | int i; | |
2116 | ||
2117 | for (i = 0; i < cachep->num; i++) { | |
b28a02de | 2118 | void *objp = slabp->s_mem + cachep->objsize * i; |
1da177e4 LT |
2119 | #if DEBUG |
2120 | /* need to poison the objs? */ | |
2121 | if (cachep->flags & SLAB_POISON) | |
2122 | poison_obj(cachep, objp, POISON_FREE); | |
2123 | if (cachep->flags & SLAB_STORE_USER) | |
2124 | *dbg_userword(cachep, objp) = NULL; | |
2125 | ||
2126 | if (cachep->flags & SLAB_RED_ZONE) { | |
2127 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; | |
2128 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
2129 | } | |
2130 | /* | |
2131 | * Constructors are not allowed to allocate memory from | |
2132 | * the same cache which they are a constructor for. | |
2133 | * Otherwise, deadlock. They must also be threaded. | |
2134 | */ | |
2135 | if (cachep->ctor && !(cachep->flags & SLAB_POISON)) | |
b28a02de PE |
2136 | cachep->ctor(objp + obj_dbghead(cachep), cachep, |
2137 | ctor_flags); | |
1da177e4 LT |
2138 | |
2139 | if (cachep->flags & SLAB_RED_ZONE) { | |
2140 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) | |
2141 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2142 | " end of an object"); |
1da177e4 LT |
2143 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) |
2144 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2145 | " start of an object"); |
1da177e4 | 2146 | } |
b28a02de PE |
2147 | if ((cachep->objsize % PAGE_SIZE) == 0 && OFF_SLAB(cachep) |
2148 | && cachep->flags & SLAB_POISON) | |
2149 | kernel_map_pages(virt_to_page(objp), | |
2150 | cachep->objsize / PAGE_SIZE, 0); | |
1da177e4 LT |
2151 | #else |
2152 | if (cachep->ctor) | |
2153 | cachep->ctor(objp, cachep, ctor_flags); | |
2154 | #endif | |
b28a02de | 2155 | slab_bufctl(slabp)[i] = i + 1; |
1da177e4 | 2156 | } |
b28a02de | 2157 | slab_bufctl(slabp)[i - 1] = BUFCTL_END; |
1da177e4 LT |
2158 | slabp->free = 0; |
2159 | } | |
2160 | ||
6daa0e28 | 2161 | static void kmem_flagcheck(kmem_cache_t *cachep, gfp_t flags) |
1da177e4 LT |
2162 | { |
2163 | if (flags & SLAB_DMA) { | |
2164 | if (!(cachep->gfpflags & GFP_DMA)) | |
2165 | BUG(); | |
2166 | } else { | |
2167 | if (cachep->gfpflags & GFP_DMA) | |
2168 | BUG(); | |
2169 | } | |
2170 | } | |
2171 | ||
2172 | static void set_slab_attr(kmem_cache_t *cachep, struct slab *slabp, void *objp) | |
2173 | { | |
2174 | int i; | |
2175 | struct page *page; | |
2176 | ||
2177 | /* Nasty!!!!!! I hope this is OK. */ | |
2178 | i = 1 << cachep->gfporder; | |
2179 | page = virt_to_page(objp); | |
2180 | do { | |
065d41cb PE |
2181 | page_set_cache(page, cachep); |
2182 | page_set_slab(page, slabp); | |
1da177e4 LT |
2183 | page++; |
2184 | } while (--i); | |
2185 | } | |
2186 | ||
2187 | /* | |
2188 | * Grow (by 1) the number of slabs within a cache. This is called by | |
2189 | * kmem_cache_alloc() when there are no active objs left in a cache. | |
2190 | */ | |
dd0fc66f | 2191 | static int cache_grow(kmem_cache_t *cachep, gfp_t flags, int nodeid) |
1da177e4 | 2192 | { |
b28a02de PE |
2193 | struct slab *slabp; |
2194 | void *objp; | |
2195 | size_t offset; | |
2196 | gfp_t local_flags; | |
2197 | unsigned long ctor_flags; | |
e498be7d | 2198 | struct kmem_list3 *l3; |
1da177e4 LT |
2199 | |
2200 | /* Be lazy and only check for valid flags here, | |
b28a02de | 2201 | * keeping it out of the critical path in kmem_cache_alloc(). |
1da177e4 | 2202 | */ |
b28a02de | 2203 | if (flags & ~(SLAB_DMA | SLAB_LEVEL_MASK | SLAB_NO_GROW)) |
1da177e4 LT |
2204 | BUG(); |
2205 | if (flags & SLAB_NO_GROW) | |
2206 | return 0; | |
2207 | ||
2208 | ctor_flags = SLAB_CTOR_CONSTRUCTOR; | |
2209 | local_flags = (flags & SLAB_LEVEL_MASK); | |
2210 | if (!(local_flags & __GFP_WAIT)) | |
2211 | /* | |
2212 | * Not allowed to sleep. Need to tell a constructor about | |
2213 | * this - it might need to know... | |
2214 | */ | |
2215 | ctor_flags |= SLAB_CTOR_ATOMIC; | |
2216 | ||
2217 | /* About to mess with non-constant members - lock. */ | |
2218 | check_irq_off(); | |
2219 | spin_lock(&cachep->spinlock); | |
2220 | ||
2221 | /* Get colour for the slab, and cal the next value. */ | |
2222 | offset = cachep->colour_next; | |
2223 | cachep->colour_next++; | |
2224 | if (cachep->colour_next >= cachep->colour) | |
2225 | cachep->colour_next = 0; | |
2226 | offset *= cachep->colour_off; | |
2227 | ||
2228 | spin_unlock(&cachep->spinlock); | |
2229 | ||
e498be7d | 2230 | check_irq_off(); |
1da177e4 LT |
2231 | if (local_flags & __GFP_WAIT) |
2232 | local_irq_enable(); | |
2233 | ||
2234 | /* | |
2235 | * The test for missing atomic flag is performed here, rather than | |
2236 | * the more obvious place, simply to reduce the critical path length | |
2237 | * in kmem_cache_alloc(). If a caller is seriously mis-behaving they | |
2238 | * will eventually be caught here (where it matters). | |
2239 | */ | |
2240 | kmem_flagcheck(cachep, flags); | |
2241 | ||
e498be7d CL |
2242 | /* Get mem for the objs. |
2243 | * Attempt to allocate a physical page from 'nodeid', | |
2244 | */ | |
1da177e4 LT |
2245 | if (!(objp = kmem_getpages(cachep, flags, nodeid))) |
2246 | goto failed; | |
2247 | ||
2248 | /* Get slab management. */ | |
2249 | if (!(slabp = alloc_slabmgmt(cachep, objp, offset, local_flags))) | |
2250 | goto opps1; | |
2251 | ||
e498be7d | 2252 | slabp->nodeid = nodeid; |
1da177e4 LT |
2253 | set_slab_attr(cachep, slabp, objp); |
2254 | ||
2255 | cache_init_objs(cachep, slabp, ctor_flags); | |
2256 | ||
2257 | if (local_flags & __GFP_WAIT) | |
2258 | local_irq_disable(); | |
2259 | check_irq_off(); | |
e498be7d CL |
2260 | l3 = cachep->nodelists[nodeid]; |
2261 | spin_lock(&l3->list_lock); | |
1da177e4 LT |
2262 | |
2263 | /* Make slab active. */ | |
e498be7d | 2264 | list_add_tail(&slabp->list, &(l3->slabs_free)); |
1da177e4 | 2265 | STATS_INC_GROWN(cachep); |
e498be7d CL |
2266 | l3->free_objects += cachep->num; |
2267 | spin_unlock(&l3->list_lock); | |
1da177e4 | 2268 | return 1; |
b28a02de | 2269 | opps1: |
1da177e4 | 2270 | kmem_freepages(cachep, objp); |
b28a02de | 2271 | failed: |
1da177e4 LT |
2272 | if (local_flags & __GFP_WAIT) |
2273 | local_irq_disable(); | |
2274 | return 0; | |
2275 | } | |
2276 | ||
2277 | #if DEBUG | |
2278 | ||
2279 | /* | |
2280 | * Perform extra freeing checks: | |
2281 | * - detect bad pointers. | |
2282 | * - POISON/RED_ZONE checking | |
2283 | * - destructor calls, for caches with POISON+dtor | |
2284 | */ | |
2285 | static void kfree_debugcheck(const void *objp) | |
2286 | { | |
2287 | struct page *page; | |
2288 | ||
2289 | if (!virt_addr_valid(objp)) { | |
2290 | printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n", | |
b28a02de PE |
2291 | (unsigned long)objp); |
2292 | BUG(); | |
1da177e4 LT |
2293 | } |
2294 | page = virt_to_page(objp); | |
2295 | if (!PageSlab(page)) { | |
b28a02de PE |
2296 | printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n", |
2297 | (unsigned long)objp); | |
1da177e4 LT |
2298 | BUG(); |
2299 | } | |
2300 | } | |
2301 | ||
2302 | static void *cache_free_debugcheck(kmem_cache_t *cachep, void *objp, | |
b28a02de | 2303 | void *caller) |
1da177e4 LT |
2304 | { |
2305 | struct page *page; | |
2306 | unsigned int objnr; | |
2307 | struct slab *slabp; | |
2308 | ||
2309 | objp -= obj_dbghead(cachep); | |
2310 | kfree_debugcheck(objp); | |
2311 | page = virt_to_page(objp); | |
2312 | ||
065d41cb | 2313 | if (page_get_cache(page) != cachep) { |
b28a02de PE |
2314 | printk(KERN_ERR |
2315 | "mismatch in kmem_cache_free: expected cache %p, got %p\n", | |
2316 | page_get_cache(page), cachep); | |
1da177e4 | 2317 | printk(KERN_ERR "%p is %s.\n", cachep, cachep->name); |
b28a02de PE |
2318 | printk(KERN_ERR "%p is %s.\n", page_get_cache(page), |
2319 | page_get_cache(page)->name); | |
1da177e4 LT |
2320 | WARN_ON(1); |
2321 | } | |
065d41cb | 2322 | slabp = page_get_slab(page); |
1da177e4 LT |
2323 | |
2324 | if (cachep->flags & SLAB_RED_ZONE) { | |
b28a02de PE |
2325 | if (*dbg_redzone1(cachep, objp) != RED_ACTIVE |
2326 | || *dbg_redzone2(cachep, objp) != RED_ACTIVE) { | |
2327 | slab_error(cachep, | |
2328 | "double free, or memory outside" | |
2329 | " object was overwritten"); | |
2330 | printk(KERN_ERR | |
2331 | "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n", | |
2332 | objp, *dbg_redzone1(cachep, objp), | |
2333 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
2334 | } |
2335 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; | |
2336 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
2337 | } | |
2338 | if (cachep->flags & SLAB_STORE_USER) | |
2339 | *dbg_userword(cachep, objp) = caller; | |
2340 | ||
b28a02de | 2341 | objnr = (objp - slabp->s_mem) / cachep->objsize; |
1da177e4 LT |
2342 | |
2343 | BUG_ON(objnr >= cachep->num); | |
b28a02de | 2344 | BUG_ON(objp != slabp->s_mem + objnr * cachep->objsize); |
1da177e4 LT |
2345 | |
2346 | if (cachep->flags & SLAB_DEBUG_INITIAL) { | |
2347 | /* Need to call the slab's constructor so the | |
2348 | * caller can perform a verify of its state (debugging). | |
2349 | * Called without the cache-lock held. | |
2350 | */ | |
b28a02de PE |
2351 | cachep->ctor(objp + obj_dbghead(cachep), |
2352 | cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY); | |
1da177e4 LT |
2353 | } |
2354 | if (cachep->flags & SLAB_POISON && cachep->dtor) { | |
2355 | /* we want to cache poison the object, | |
2356 | * call the destruction callback | |
2357 | */ | |
b28a02de | 2358 | cachep->dtor(objp + obj_dbghead(cachep), cachep, 0); |
1da177e4 LT |
2359 | } |
2360 | if (cachep->flags & SLAB_POISON) { | |
2361 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
2362 | if ((cachep->objsize % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) { | |
2363 | store_stackinfo(cachep, objp, (unsigned long)caller); | |
b28a02de PE |
2364 | kernel_map_pages(virt_to_page(objp), |
2365 | cachep->objsize / PAGE_SIZE, 0); | |
1da177e4 LT |
2366 | } else { |
2367 | poison_obj(cachep, objp, POISON_FREE); | |
2368 | } | |
2369 | #else | |
2370 | poison_obj(cachep, objp, POISON_FREE); | |
2371 | #endif | |
2372 | } | |
2373 | return objp; | |
2374 | } | |
2375 | ||
2376 | static void check_slabp(kmem_cache_t *cachep, struct slab *slabp) | |
2377 | { | |
2378 | kmem_bufctl_t i; | |
2379 | int entries = 0; | |
b28a02de | 2380 | |
1da177e4 LT |
2381 | /* Check slab's freelist to see if this obj is there. */ |
2382 | for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) { | |
2383 | entries++; | |
2384 | if (entries > cachep->num || i >= cachep->num) | |
2385 | goto bad; | |
2386 | } | |
2387 | if (entries != cachep->num - slabp->inuse) { | |
b28a02de PE |
2388 | bad: |
2389 | printk(KERN_ERR | |
2390 | "slab: Internal list corruption detected in cache '%s'(%d), slabp %p(%d). Hexdump:\n", | |
2391 | cachep->name, cachep->num, slabp, slabp->inuse); | |
2392 | for (i = 0; | |
2393 | i < sizeof(slabp) + cachep->num * sizeof(kmem_bufctl_t); | |
2394 | i++) { | |
2395 | if ((i % 16) == 0) | |
1da177e4 | 2396 | printk("\n%03x:", i); |
b28a02de | 2397 | printk(" %02x", ((unsigned char *)slabp)[i]); |
1da177e4 LT |
2398 | } |
2399 | printk("\n"); | |
2400 | BUG(); | |
2401 | } | |
2402 | } | |
2403 | #else | |
2404 | #define kfree_debugcheck(x) do { } while(0) | |
2405 | #define cache_free_debugcheck(x,objp,z) (objp) | |
2406 | #define check_slabp(x,y) do { } while(0) | |
2407 | #endif | |
2408 | ||
dd0fc66f | 2409 | static void *cache_alloc_refill(kmem_cache_t *cachep, gfp_t flags) |
1da177e4 LT |
2410 | { |
2411 | int batchcount; | |
2412 | struct kmem_list3 *l3; | |
2413 | struct array_cache *ac; | |
2414 | ||
2415 | check_irq_off(); | |
2416 | ac = ac_data(cachep); | |
b28a02de | 2417 | retry: |
1da177e4 LT |
2418 | batchcount = ac->batchcount; |
2419 | if (!ac->touched && batchcount > BATCHREFILL_LIMIT) { | |
2420 | /* if there was little recent activity on this | |
2421 | * cache, then perform only a partial refill. | |
2422 | * Otherwise we could generate refill bouncing. | |
2423 | */ | |
2424 | batchcount = BATCHREFILL_LIMIT; | |
2425 | } | |
e498be7d CL |
2426 | l3 = cachep->nodelists[numa_node_id()]; |
2427 | ||
2428 | BUG_ON(ac->avail > 0 || !l3); | |
2429 | spin_lock(&l3->list_lock); | |
1da177e4 | 2430 | |
1da177e4 LT |
2431 | if (l3->shared) { |
2432 | struct array_cache *shared_array = l3->shared; | |
2433 | if (shared_array->avail) { | |
2434 | if (batchcount > shared_array->avail) | |
2435 | batchcount = shared_array->avail; | |
2436 | shared_array->avail -= batchcount; | |
2437 | ac->avail = batchcount; | |
e498be7d | 2438 | memcpy(ac->entry, |
b28a02de PE |
2439 | &(shared_array->entry[shared_array->avail]), |
2440 | sizeof(void *) * batchcount); | |
1da177e4 LT |
2441 | shared_array->touched = 1; |
2442 | goto alloc_done; | |
2443 | } | |
2444 | } | |
2445 | while (batchcount > 0) { | |
2446 | struct list_head *entry; | |
2447 | struct slab *slabp; | |
2448 | /* Get slab alloc is to come from. */ | |
2449 | entry = l3->slabs_partial.next; | |
2450 | if (entry == &l3->slabs_partial) { | |
2451 | l3->free_touched = 1; | |
2452 | entry = l3->slabs_free.next; | |
2453 | if (entry == &l3->slabs_free) | |
2454 | goto must_grow; | |
2455 | } | |
2456 | ||
2457 | slabp = list_entry(entry, struct slab, list); | |
2458 | check_slabp(cachep, slabp); | |
2459 | check_spinlock_acquired(cachep); | |
2460 | while (slabp->inuse < cachep->num && batchcount--) { | |
2461 | kmem_bufctl_t next; | |
2462 | STATS_INC_ALLOCED(cachep); | |
2463 | STATS_INC_ACTIVE(cachep); | |
2464 | STATS_SET_HIGH(cachep); | |
2465 | ||
2466 | /* get obj pointer */ | |
e498be7d | 2467 | ac->entry[ac->avail++] = slabp->s_mem + |
b28a02de | 2468 | slabp->free * cachep->objsize; |
1da177e4 LT |
2469 | |
2470 | slabp->inuse++; | |
2471 | next = slab_bufctl(slabp)[slabp->free]; | |
2472 | #if DEBUG | |
2473 | slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE; | |
09ad4bbc | 2474 | WARN_ON(numa_node_id() != slabp->nodeid); |
1da177e4 | 2475 | #endif |
b28a02de | 2476 | slabp->free = next; |
1da177e4 LT |
2477 | } |
2478 | check_slabp(cachep, slabp); | |
2479 | ||
2480 | /* move slabp to correct slabp list: */ | |
2481 | list_del(&slabp->list); | |
2482 | if (slabp->free == BUFCTL_END) | |
2483 | list_add(&slabp->list, &l3->slabs_full); | |
2484 | else | |
2485 | list_add(&slabp->list, &l3->slabs_partial); | |
2486 | } | |
2487 | ||
b28a02de | 2488 | must_grow: |
1da177e4 | 2489 | l3->free_objects -= ac->avail; |
b28a02de | 2490 | alloc_done: |
e498be7d | 2491 | spin_unlock(&l3->list_lock); |
1da177e4 LT |
2492 | |
2493 | if (unlikely(!ac->avail)) { | |
2494 | int x; | |
e498be7d CL |
2495 | x = cache_grow(cachep, flags, numa_node_id()); |
2496 | ||
1da177e4 LT |
2497 | // cache_grow can reenable interrupts, then ac could change. |
2498 | ac = ac_data(cachep); | |
2499 | if (!x && ac->avail == 0) // no objects in sight? abort | |
2500 | return NULL; | |
2501 | ||
b28a02de | 2502 | if (!ac->avail) // objects refilled by interrupt? |
1da177e4 LT |
2503 | goto retry; |
2504 | } | |
2505 | ac->touched = 1; | |
e498be7d | 2506 | return ac->entry[--ac->avail]; |
1da177e4 LT |
2507 | } |
2508 | ||
2509 | static inline void | |
dd0fc66f | 2510 | cache_alloc_debugcheck_before(kmem_cache_t *cachep, gfp_t flags) |
1da177e4 LT |
2511 | { |
2512 | might_sleep_if(flags & __GFP_WAIT); | |
2513 | #if DEBUG | |
2514 | kmem_flagcheck(cachep, flags); | |
2515 | #endif | |
2516 | } | |
2517 | ||
2518 | #if DEBUG | |
b28a02de PE |
2519 | static void *cache_alloc_debugcheck_after(kmem_cache_t *cachep, gfp_t flags, |
2520 | void *objp, void *caller) | |
1da177e4 | 2521 | { |
b28a02de | 2522 | if (!objp) |
1da177e4 | 2523 | return objp; |
b28a02de | 2524 | if (cachep->flags & SLAB_POISON) { |
1da177e4 LT |
2525 | #ifdef CONFIG_DEBUG_PAGEALLOC |
2526 | if ((cachep->objsize % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) | |
b28a02de PE |
2527 | kernel_map_pages(virt_to_page(objp), |
2528 | cachep->objsize / PAGE_SIZE, 1); | |
1da177e4 LT |
2529 | else |
2530 | check_poison_obj(cachep, objp); | |
2531 | #else | |
2532 | check_poison_obj(cachep, objp); | |
2533 | #endif | |
2534 | poison_obj(cachep, objp, POISON_INUSE); | |
2535 | } | |
2536 | if (cachep->flags & SLAB_STORE_USER) | |
2537 | *dbg_userword(cachep, objp) = caller; | |
2538 | ||
2539 | if (cachep->flags & SLAB_RED_ZONE) { | |
b28a02de PE |
2540 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE |
2541 | || *dbg_redzone2(cachep, objp) != RED_INACTIVE) { | |
2542 | slab_error(cachep, | |
2543 | "double free, or memory outside" | |
2544 | " object was overwritten"); | |
2545 | printk(KERN_ERR | |
2546 | "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n", | |
2547 | objp, *dbg_redzone1(cachep, objp), | |
2548 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
2549 | } |
2550 | *dbg_redzone1(cachep, objp) = RED_ACTIVE; | |
2551 | *dbg_redzone2(cachep, objp) = RED_ACTIVE; | |
2552 | } | |
2553 | objp += obj_dbghead(cachep); | |
2554 | if (cachep->ctor && cachep->flags & SLAB_POISON) { | |
b28a02de | 2555 | unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR; |
1da177e4 LT |
2556 | |
2557 | if (!(flags & __GFP_WAIT)) | |
2558 | ctor_flags |= SLAB_CTOR_ATOMIC; | |
2559 | ||
2560 | cachep->ctor(objp, cachep, ctor_flags); | |
b28a02de | 2561 | } |
1da177e4 LT |
2562 | return objp; |
2563 | } | |
2564 | #else | |
2565 | #define cache_alloc_debugcheck_after(a,b,objp,d) (objp) | |
2566 | #endif | |
2567 | ||
dd0fc66f | 2568 | static inline void *____cache_alloc(kmem_cache_t *cachep, gfp_t flags) |
1da177e4 | 2569 | { |
b28a02de | 2570 | void *objp; |
1da177e4 LT |
2571 | struct array_cache *ac; |
2572 | ||
5c382300 | 2573 | check_irq_off(); |
1da177e4 LT |
2574 | ac = ac_data(cachep); |
2575 | if (likely(ac->avail)) { | |
2576 | STATS_INC_ALLOCHIT(cachep); | |
2577 | ac->touched = 1; | |
e498be7d | 2578 | objp = ac->entry[--ac->avail]; |
1da177e4 LT |
2579 | } else { |
2580 | STATS_INC_ALLOCMISS(cachep); | |
2581 | objp = cache_alloc_refill(cachep, flags); | |
2582 | } | |
5c382300 AK |
2583 | return objp; |
2584 | } | |
2585 | ||
dd0fc66f | 2586 | static inline void *__cache_alloc(kmem_cache_t *cachep, gfp_t flags) |
5c382300 AK |
2587 | { |
2588 | unsigned long save_flags; | |
b28a02de | 2589 | void *objp; |
5c382300 AK |
2590 | |
2591 | cache_alloc_debugcheck_before(cachep, flags); | |
2592 | ||
2593 | local_irq_save(save_flags); | |
2594 | objp = ____cache_alloc(cachep, flags); | |
1da177e4 | 2595 | local_irq_restore(save_flags); |
34342e86 | 2596 | objp = cache_alloc_debugcheck_after(cachep, flags, objp, |
b28a02de | 2597 | __builtin_return_address(0)); |
34342e86 | 2598 | prefetchw(objp); |
1da177e4 LT |
2599 | return objp; |
2600 | } | |
2601 | ||
e498be7d CL |
2602 | #ifdef CONFIG_NUMA |
2603 | /* | |
2604 | * A interface to enable slab creation on nodeid | |
1da177e4 | 2605 | */ |
6daa0e28 | 2606 | static void *__cache_alloc_node(kmem_cache_t *cachep, gfp_t flags, int nodeid) |
e498be7d CL |
2607 | { |
2608 | struct list_head *entry; | |
b28a02de PE |
2609 | struct slab *slabp; |
2610 | struct kmem_list3 *l3; | |
2611 | void *obj; | |
2612 | kmem_bufctl_t next; | |
2613 | int x; | |
2614 | ||
2615 | l3 = cachep->nodelists[nodeid]; | |
2616 | BUG_ON(!l3); | |
2617 | ||
2618 | retry: | |
2619 | spin_lock(&l3->list_lock); | |
2620 | entry = l3->slabs_partial.next; | |
2621 | if (entry == &l3->slabs_partial) { | |
2622 | l3->free_touched = 1; | |
2623 | entry = l3->slabs_free.next; | |
2624 | if (entry == &l3->slabs_free) | |
2625 | goto must_grow; | |
2626 | } | |
2627 | ||
2628 | slabp = list_entry(entry, struct slab, list); | |
2629 | check_spinlock_acquired_node(cachep, nodeid); | |
2630 | check_slabp(cachep, slabp); | |
2631 | ||
2632 | STATS_INC_NODEALLOCS(cachep); | |
2633 | STATS_INC_ACTIVE(cachep); | |
2634 | STATS_SET_HIGH(cachep); | |
2635 | ||
2636 | BUG_ON(slabp->inuse == cachep->num); | |
2637 | ||
2638 | /* get obj pointer */ | |
2639 | obj = slabp->s_mem + slabp->free * cachep->objsize; | |
2640 | slabp->inuse++; | |
2641 | next = slab_bufctl(slabp)[slabp->free]; | |
e498be7d | 2642 | #if DEBUG |
b28a02de | 2643 | slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE; |
e498be7d | 2644 | #endif |
b28a02de PE |
2645 | slabp->free = next; |
2646 | check_slabp(cachep, slabp); | |
2647 | l3->free_objects--; | |
2648 | /* move slabp to correct slabp list: */ | |
2649 | list_del(&slabp->list); | |
2650 | ||
2651 | if (slabp->free == BUFCTL_END) { | |
2652 | list_add(&slabp->list, &l3->slabs_full); | |
2653 | } else { | |
2654 | list_add(&slabp->list, &l3->slabs_partial); | |
2655 | } | |
e498be7d | 2656 | |
b28a02de PE |
2657 | spin_unlock(&l3->list_lock); |
2658 | goto done; | |
e498be7d | 2659 | |
b28a02de PE |
2660 | must_grow: |
2661 | spin_unlock(&l3->list_lock); | |
2662 | x = cache_grow(cachep, flags, nodeid); | |
1da177e4 | 2663 | |
b28a02de PE |
2664 | if (!x) |
2665 | return NULL; | |
e498be7d | 2666 | |
b28a02de PE |
2667 | goto retry; |
2668 | done: | |
2669 | return obj; | |
e498be7d CL |
2670 | } |
2671 | #endif | |
2672 | ||
2673 | /* | |
2674 | * Caller needs to acquire correct kmem_list's list_lock | |
2675 | */ | |
b28a02de PE |
2676 | static void free_block(kmem_cache_t *cachep, void **objpp, int nr_objects, |
2677 | int node) | |
1da177e4 LT |
2678 | { |
2679 | int i; | |
e498be7d | 2680 | struct kmem_list3 *l3; |
1da177e4 LT |
2681 | |
2682 | for (i = 0; i < nr_objects; i++) { | |
2683 | void *objp = objpp[i]; | |
2684 | struct slab *slabp; | |
2685 | unsigned int objnr; | |
2686 | ||
065d41cb | 2687 | slabp = page_get_slab(virt_to_page(objp)); |
ff69416e | 2688 | l3 = cachep->nodelists[node]; |
1da177e4 LT |
2689 | list_del(&slabp->list); |
2690 | objnr = (objp - slabp->s_mem) / cachep->objsize; | |
ff69416e | 2691 | check_spinlock_acquired_node(cachep, node); |
1da177e4 | 2692 | check_slabp(cachep, slabp); |
e498be7d | 2693 | |
1da177e4 | 2694 | #if DEBUG |
09ad4bbc CL |
2695 | /* Verify that the slab belongs to the intended node */ |
2696 | WARN_ON(slabp->nodeid != node); | |
2697 | ||
1da177e4 | 2698 | if (slab_bufctl(slabp)[objnr] != BUFCTL_FREE) { |
e498be7d | 2699 | printk(KERN_ERR "slab: double free detected in cache " |
b28a02de | 2700 | "'%s', objp %p\n", cachep->name, objp); |
1da177e4 LT |
2701 | BUG(); |
2702 | } | |
2703 | #endif | |
2704 | slab_bufctl(slabp)[objnr] = slabp->free; | |
2705 | slabp->free = objnr; | |
2706 | STATS_DEC_ACTIVE(cachep); | |
2707 | slabp->inuse--; | |
e498be7d | 2708 | l3->free_objects++; |
1da177e4 LT |
2709 | check_slabp(cachep, slabp); |
2710 | ||
2711 | /* fixup slab chains */ | |
2712 | if (slabp->inuse == 0) { | |
e498be7d CL |
2713 | if (l3->free_objects > l3->free_limit) { |
2714 | l3->free_objects -= cachep->num; | |
1da177e4 LT |
2715 | slab_destroy(cachep, slabp); |
2716 | } else { | |
e498be7d | 2717 | list_add(&slabp->list, &l3->slabs_free); |
1da177e4 LT |
2718 | } |
2719 | } else { | |
2720 | /* Unconditionally move a slab to the end of the | |
2721 | * partial list on free - maximum time for the | |
2722 | * other objects to be freed, too. | |
2723 | */ | |
e498be7d | 2724 | list_add_tail(&slabp->list, &l3->slabs_partial); |
1da177e4 LT |
2725 | } |
2726 | } | |
2727 | } | |
2728 | ||
2729 | static void cache_flusharray(kmem_cache_t *cachep, struct array_cache *ac) | |
2730 | { | |
2731 | int batchcount; | |
e498be7d | 2732 | struct kmem_list3 *l3; |
ff69416e | 2733 | int node = numa_node_id(); |
1da177e4 LT |
2734 | |
2735 | batchcount = ac->batchcount; | |
2736 | #if DEBUG | |
2737 | BUG_ON(!batchcount || batchcount > ac->avail); | |
2738 | #endif | |
2739 | check_irq_off(); | |
ff69416e | 2740 | l3 = cachep->nodelists[node]; |
e498be7d CL |
2741 | spin_lock(&l3->list_lock); |
2742 | if (l3->shared) { | |
2743 | struct array_cache *shared_array = l3->shared; | |
b28a02de | 2744 | int max = shared_array->limit - shared_array->avail; |
1da177e4 LT |
2745 | if (max) { |
2746 | if (batchcount > max) | |
2747 | batchcount = max; | |
e498be7d | 2748 | memcpy(&(shared_array->entry[shared_array->avail]), |
b28a02de | 2749 | ac->entry, sizeof(void *) * batchcount); |
1da177e4 LT |
2750 | shared_array->avail += batchcount; |
2751 | goto free_done; | |
2752 | } | |
2753 | } | |
2754 | ||
ff69416e | 2755 | free_block(cachep, ac->entry, batchcount, node); |
b28a02de | 2756 | free_done: |
1da177e4 LT |
2757 | #if STATS |
2758 | { | |
2759 | int i = 0; | |
2760 | struct list_head *p; | |
2761 | ||
e498be7d CL |
2762 | p = l3->slabs_free.next; |
2763 | while (p != &(l3->slabs_free)) { | |
1da177e4 LT |
2764 | struct slab *slabp; |
2765 | ||
2766 | slabp = list_entry(p, struct slab, list); | |
2767 | BUG_ON(slabp->inuse); | |
2768 | ||
2769 | i++; | |
2770 | p = p->next; | |
2771 | } | |
2772 | STATS_SET_FREEABLE(cachep, i); | |
2773 | } | |
2774 | #endif | |
e498be7d | 2775 | spin_unlock(&l3->list_lock); |
1da177e4 | 2776 | ac->avail -= batchcount; |
e498be7d | 2777 | memmove(ac->entry, &(ac->entry[batchcount]), |
b28a02de | 2778 | sizeof(void *) * ac->avail); |
1da177e4 LT |
2779 | } |
2780 | ||
2781 | /* | |
2782 | * __cache_free | |
2783 | * Release an obj back to its cache. If the obj has a constructed | |
2784 | * state, it must be in this state _before_ it is released. | |
2785 | * | |
2786 | * Called with disabled ints. | |
2787 | */ | |
2788 | static inline void __cache_free(kmem_cache_t *cachep, void *objp) | |
2789 | { | |
2790 | struct array_cache *ac = ac_data(cachep); | |
2791 | ||
2792 | check_irq_off(); | |
2793 | objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0)); | |
2794 | ||
e498be7d CL |
2795 | /* Make sure we are not freeing a object from another |
2796 | * node to the array cache on this cpu. | |
2797 | */ | |
2798 | #ifdef CONFIG_NUMA | |
2799 | { | |
2800 | struct slab *slabp; | |
065d41cb | 2801 | slabp = page_get_slab(virt_to_page(objp)); |
e498be7d CL |
2802 | if (unlikely(slabp->nodeid != numa_node_id())) { |
2803 | struct array_cache *alien = NULL; | |
2804 | int nodeid = slabp->nodeid; | |
b28a02de PE |
2805 | struct kmem_list3 *l3 = |
2806 | cachep->nodelists[numa_node_id()]; | |
e498be7d CL |
2807 | |
2808 | STATS_INC_NODEFREES(cachep); | |
2809 | if (l3->alien && l3->alien[nodeid]) { | |
2810 | alien = l3->alien[nodeid]; | |
2811 | spin_lock(&alien->lock); | |
2812 | if (unlikely(alien->avail == alien->limit)) | |
2813 | __drain_alien_cache(cachep, | |
b28a02de | 2814 | alien, nodeid); |
e498be7d CL |
2815 | alien->entry[alien->avail++] = objp; |
2816 | spin_unlock(&alien->lock); | |
2817 | } else { | |
2818 | spin_lock(&(cachep->nodelists[nodeid])-> | |
b28a02de | 2819 | list_lock); |
ff69416e | 2820 | free_block(cachep, &objp, 1, nodeid); |
e498be7d | 2821 | spin_unlock(&(cachep->nodelists[nodeid])-> |
b28a02de | 2822 | list_lock); |
e498be7d CL |
2823 | } |
2824 | return; | |
2825 | } | |
2826 | } | |
2827 | #endif | |
1da177e4 LT |
2828 | if (likely(ac->avail < ac->limit)) { |
2829 | STATS_INC_FREEHIT(cachep); | |
e498be7d | 2830 | ac->entry[ac->avail++] = objp; |
1da177e4 LT |
2831 | return; |
2832 | } else { | |
2833 | STATS_INC_FREEMISS(cachep); | |
2834 | cache_flusharray(cachep, ac); | |
e498be7d | 2835 | ac->entry[ac->avail++] = objp; |
1da177e4 LT |
2836 | } |
2837 | } | |
2838 | ||
2839 | /** | |
2840 | * kmem_cache_alloc - Allocate an object | |
2841 | * @cachep: The cache to allocate from. | |
2842 | * @flags: See kmalloc(). | |
2843 | * | |
2844 | * Allocate an object from this cache. The flags are only relevant | |
2845 | * if the cache has no available objects. | |
2846 | */ | |
dd0fc66f | 2847 | void *kmem_cache_alloc(kmem_cache_t *cachep, gfp_t flags) |
1da177e4 LT |
2848 | { |
2849 | return __cache_alloc(cachep, flags); | |
2850 | } | |
2851 | EXPORT_SYMBOL(kmem_cache_alloc); | |
2852 | ||
2853 | /** | |
2854 | * kmem_ptr_validate - check if an untrusted pointer might | |
2855 | * be a slab entry. | |
2856 | * @cachep: the cache we're checking against | |
2857 | * @ptr: pointer to validate | |
2858 | * | |
2859 | * This verifies that the untrusted pointer looks sane: | |
2860 | * it is _not_ a guarantee that the pointer is actually | |
2861 | * part of the slab cache in question, but it at least | |
2862 | * validates that the pointer can be dereferenced and | |
2863 | * looks half-way sane. | |
2864 | * | |
2865 | * Currently only used for dentry validation. | |
2866 | */ | |
2867 | int fastcall kmem_ptr_validate(kmem_cache_t *cachep, void *ptr) | |
2868 | { | |
b28a02de | 2869 | unsigned long addr = (unsigned long)ptr; |
1da177e4 | 2870 | unsigned long min_addr = PAGE_OFFSET; |
b28a02de | 2871 | unsigned long align_mask = BYTES_PER_WORD - 1; |
1da177e4 LT |
2872 | unsigned long size = cachep->objsize; |
2873 | struct page *page; | |
2874 | ||
2875 | if (unlikely(addr < min_addr)) | |
2876 | goto out; | |
2877 | if (unlikely(addr > (unsigned long)high_memory - size)) | |
2878 | goto out; | |
2879 | if (unlikely(addr & align_mask)) | |
2880 | goto out; | |
2881 | if (unlikely(!kern_addr_valid(addr))) | |
2882 | goto out; | |
2883 | if (unlikely(!kern_addr_valid(addr + size - 1))) | |
2884 | goto out; | |
2885 | page = virt_to_page(ptr); | |
2886 | if (unlikely(!PageSlab(page))) | |
2887 | goto out; | |
065d41cb | 2888 | if (unlikely(page_get_cache(page) != cachep)) |
1da177e4 LT |
2889 | goto out; |
2890 | return 1; | |
b28a02de | 2891 | out: |
1da177e4 LT |
2892 | return 0; |
2893 | } | |
2894 | ||
2895 | #ifdef CONFIG_NUMA | |
2896 | /** | |
2897 | * kmem_cache_alloc_node - Allocate an object on the specified node | |
2898 | * @cachep: The cache to allocate from. | |
2899 | * @flags: See kmalloc(). | |
2900 | * @nodeid: node number of the target node. | |
2901 | * | |
2902 | * Identical to kmem_cache_alloc, except that this function is slow | |
2903 | * and can sleep. And it will allocate memory on the given node, which | |
2904 | * can improve the performance for cpu bound structures. | |
e498be7d CL |
2905 | * New and improved: it will now make sure that the object gets |
2906 | * put on the correct node list so that there is no false sharing. | |
1da177e4 | 2907 | */ |
dd0fc66f | 2908 | void *kmem_cache_alloc_node(kmem_cache_t *cachep, gfp_t flags, int nodeid) |
1da177e4 | 2909 | { |
e498be7d CL |
2910 | unsigned long save_flags; |
2911 | void *ptr; | |
1da177e4 | 2912 | |
ff69416e | 2913 | if (nodeid == -1) |
e498be7d | 2914 | return __cache_alloc(cachep, flags); |
1da177e4 | 2915 | |
e498be7d CL |
2916 | if (unlikely(!cachep->nodelists[nodeid])) { |
2917 | /* Fall back to __cache_alloc if we run into trouble */ | |
b28a02de PE |
2918 | printk(KERN_WARNING |
2919 | "slab: not allocating in inactive node %d for cache %s\n", | |
2920 | nodeid, cachep->name); | |
2921 | return __cache_alloc(cachep, flags); | |
1da177e4 | 2922 | } |
1da177e4 | 2923 | |
e498be7d CL |
2924 | cache_alloc_debugcheck_before(cachep, flags); |
2925 | local_irq_save(save_flags); | |
5c382300 AK |
2926 | if (nodeid == numa_node_id()) |
2927 | ptr = ____cache_alloc(cachep, flags); | |
2928 | else | |
2929 | ptr = __cache_alloc_node(cachep, flags, nodeid); | |
e498be7d | 2930 | local_irq_restore(save_flags); |
b28a02de PE |
2931 | ptr = |
2932 | cache_alloc_debugcheck_after(cachep, flags, ptr, | |
2933 | __builtin_return_address(0)); | |
1da177e4 | 2934 | |
e498be7d | 2935 | return ptr; |
1da177e4 LT |
2936 | } |
2937 | EXPORT_SYMBOL(kmem_cache_alloc_node); | |
2938 | ||
dd0fc66f | 2939 | void *kmalloc_node(size_t size, gfp_t flags, int node) |
97e2bde4 MS |
2940 | { |
2941 | kmem_cache_t *cachep; | |
2942 | ||
2943 | cachep = kmem_find_general_cachep(size, flags); | |
2944 | if (unlikely(cachep == NULL)) | |
2945 | return NULL; | |
2946 | return kmem_cache_alloc_node(cachep, flags, node); | |
2947 | } | |
2948 | EXPORT_SYMBOL(kmalloc_node); | |
1da177e4 LT |
2949 | #endif |
2950 | ||
2951 | /** | |
2952 | * kmalloc - allocate memory | |
2953 | * @size: how many bytes of memory are required. | |
2954 | * @flags: the type of memory to allocate. | |
2955 | * | |
2956 | * kmalloc is the normal method of allocating memory | |
2957 | * in the kernel. | |
2958 | * | |
2959 | * The @flags argument may be one of: | |
2960 | * | |
2961 | * %GFP_USER - Allocate memory on behalf of user. May sleep. | |
2962 | * | |
2963 | * %GFP_KERNEL - Allocate normal kernel ram. May sleep. | |
2964 | * | |
2965 | * %GFP_ATOMIC - Allocation will not sleep. Use inside interrupt handlers. | |
2966 | * | |
2967 | * Additionally, the %GFP_DMA flag may be set to indicate the memory | |
2968 | * must be suitable for DMA. This can mean different things on different | |
2969 | * platforms. For example, on i386, it means that the memory must come | |
2970 | * from the first 16MB. | |
2971 | */ | |
dd0fc66f | 2972 | void *__kmalloc(size_t size, gfp_t flags) |
1da177e4 LT |
2973 | { |
2974 | kmem_cache_t *cachep; | |
2975 | ||
97e2bde4 MS |
2976 | /* If you want to save a few bytes .text space: replace |
2977 | * __ with kmem_. | |
2978 | * Then kmalloc uses the uninlined functions instead of the inline | |
2979 | * functions. | |
2980 | */ | |
2981 | cachep = __find_general_cachep(size, flags); | |
dbdb9045 AM |
2982 | if (unlikely(cachep == NULL)) |
2983 | return NULL; | |
1da177e4 LT |
2984 | return __cache_alloc(cachep, flags); |
2985 | } | |
2986 | EXPORT_SYMBOL(__kmalloc); | |
2987 | ||
2988 | #ifdef CONFIG_SMP | |
2989 | /** | |
2990 | * __alloc_percpu - allocate one copy of the object for every present | |
2991 | * cpu in the system, zeroing them. | |
2992 | * Objects should be dereferenced using the per_cpu_ptr macro only. | |
2993 | * | |
2994 | * @size: how many bytes of memory are required. | |
1da177e4 | 2995 | */ |
f9f75005 | 2996 | void *__alloc_percpu(size_t size) |
1da177e4 LT |
2997 | { |
2998 | int i; | |
b28a02de | 2999 | struct percpu_data *pdata = kmalloc(sizeof(*pdata), GFP_KERNEL); |
1da177e4 LT |
3000 | |
3001 | if (!pdata) | |
3002 | return NULL; | |
3003 | ||
e498be7d CL |
3004 | /* |
3005 | * Cannot use for_each_online_cpu since a cpu may come online | |
3006 | * and we have no way of figuring out how to fix the array | |
3007 | * that we have allocated then.... | |
3008 | */ | |
3009 | for_each_cpu(i) { | |
3010 | int node = cpu_to_node(i); | |
3011 | ||
3012 | if (node_online(node)) | |
3013 | pdata->ptrs[i] = kmalloc_node(size, GFP_KERNEL, node); | |
3014 | else | |
3015 | pdata->ptrs[i] = kmalloc(size, GFP_KERNEL); | |
1da177e4 LT |
3016 | |
3017 | if (!pdata->ptrs[i]) | |
3018 | goto unwind_oom; | |
3019 | memset(pdata->ptrs[i], 0, size); | |
3020 | } | |
3021 | ||
3022 | /* Catch derefs w/o wrappers */ | |
b28a02de | 3023 | return (void *)(~(unsigned long)pdata); |
1da177e4 | 3024 | |
b28a02de | 3025 | unwind_oom: |
1da177e4 LT |
3026 | while (--i >= 0) { |
3027 | if (!cpu_possible(i)) | |
3028 | continue; | |
3029 | kfree(pdata->ptrs[i]); | |
3030 | } | |
3031 | kfree(pdata); | |
3032 | return NULL; | |
3033 | } | |
3034 | EXPORT_SYMBOL(__alloc_percpu); | |
3035 | #endif | |
3036 | ||
3037 | /** | |
3038 | * kmem_cache_free - Deallocate an object | |
3039 | * @cachep: The cache the allocation was from. | |
3040 | * @objp: The previously allocated object. | |
3041 | * | |
3042 | * Free an object which was previously allocated from this | |
3043 | * cache. | |
3044 | */ | |
3045 | void kmem_cache_free(kmem_cache_t *cachep, void *objp) | |
3046 | { | |
3047 | unsigned long flags; | |
3048 | ||
3049 | local_irq_save(flags); | |
3050 | __cache_free(cachep, objp); | |
3051 | local_irq_restore(flags); | |
3052 | } | |
3053 | EXPORT_SYMBOL(kmem_cache_free); | |
3054 | ||
3055 | /** | |
dd392710 PE |
3056 | * kzalloc - allocate memory. The memory is set to zero. |
3057 | * @size: how many bytes of memory are required. | |
1da177e4 LT |
3058 | * @flags: the type of memory to allocate. |
3059 | */ | |
dd0fc66f | 3060 | void *kzalloc(size_t size, gfp_t flags) |
1da177e4 | 3061 | { |
dd392710 | 3062 | void *ret = kmalloc(size, flags); |
1da177e4 | 3063 | if (ret) |
dd392710 | 3064 | memset(ret, 0, size); |
1da177e4 LT |
3065 | return ret; |
3066 | } | |
dd392710 | 3067 | EXPORT_SYMBOL(kzalloc); |
1da177e4 LT |
3068 | |
3069 | /** | |
3070 | * kfree - free previously allocated memory | |
3071 | * @objp: pointer returned by kmalloc. | |
3072 | * | |
80e93eff PE |
3073 | * If @objp is NULL, no operation is performed. |
3074 | * | |
1da177e4 LT |
3075 | * Don't free memory not originally allocated by kmalloc() |
3076 | * or you will run into trouble. | |
3077 | */ | |
3078 | void kfree(const void *objp) | |
3079 | { | |
3080 | kmem_cache_t *c; | |
3081 | unsigned long flags; | |
3082 | ||
3083 | if (unlikely(!objp)) | |
3084 | return; | |
3085 | local_irq_save(flags); | |
3086 | kfree_debugcheck(objp); | |
065d41cb | 3087 | c = page_get_cache(virt_to_page(objp)); |
b28a02de | 3088 | __cache_free(c, (void *)objp); |
1da177e4 LT |
3089 | local_irq_restore(flags); |
3090 | } | |
3091 | EXPORT_SYMBOL(kfree); | |
3092 | ||
3093 | #ifdef CONFIG_SMP | |
3094 | /** | |
3095 | * free_percpu - free previously allocated percpu memory | |
3096 | * @objp: pointer returned by alloc_percpu. | |
3097 | * | |
3098 | * Don't free memory not originally allocated by alloc_percpu() | |
3099 | * The complemented objp is to check for that. | |
3100 | */ | |
b28a02de | 3101 | void free_percpu(const void *objp) |
1da177e4 LT |
3102 | { |
3103 | int i; | |
b28a02de | 3104 | struct percpu_data *p = (struct percpu_data *)(~(unsigned long)objp); |
1da177e4 | 3105 | |
e498be7d CL |
3106 | /* |
3107 | * We allocate for all cpus so we cannot use for online cpu here. | |
3108 | */ | |
3109 | for_each_cpu(i) | |
b28a02de | 3110 | kfree(p->ptrs[i]); |
1da177e4 LT |
3111 | kfree(p); |
3112 | } | |
3113 | EXPORT_SYMBOL(free_percpu); | |
3114 | #endif | |
3115 | ||
3116 | unsigned int kmem_cache_size(kmem_cache_t *cachep) | |
3117 | { | |
3118 | return obj_reallen(cachep); | |
3119 | } | |
3120 | EXPORT_SYMBOL(kmem_cache_size); | |
3121 | ||
1944972d ACM |
3122 | const char *kmem_cache_name(kmem_cache_t *cachep) |
3123 | { | |
3124 | return cachep->name; | |
3125 | } | |
3126 | EXPORT_SYMBOL_GPL(kmem_cache_name); | |
3127 | ||
e498be7d CL |
3128 | /* |
3129 | * This initializes kmem_list3 for all nodes. | |
3130 | */ | |
3131 | static int alloc_kmemlist(kmem_cache_t *cachep) | |
3132 | { | |
3133 | int node; | |
3134 | struct kmem_list3 *l3; | |
3135 | int err = 0; | |
3136 | ||
3137 | for_each_online_node(node) { | |
3138 | struct array_cache *nc = NULL, *new; | |
3139 | struct array_cache **new_alien = NULL; | |
3140 | #ifdef CONFIG_NUMA | |
3141 | if (!(new_alien = alloc_alien_cache(node, cachep->limit))) | |
3142 | goto fail; | |
3143 | #endif | |
b28a02de PE |
3144 | if (!(new = alloc_arraycache(node, (cachep->shared * |
3145 | cachep->batchcount), | |
3146 | 0xbaadf00d))) | |
e498be7d CL |
3147 | goto fail; |
3148 | if ((l3 = cachep->nodelists[node])) { | |
3149 | ||
3150 | spin_lock_irq(&l3->list_lock); | |
3151 | ||
3152 | if ((nc = cachep->nodelists[node]->shared)) | |
b28a02de | 3153 | free_block(cachep, nc->entry, nc->avail, node); |
e498be7d CL |
3154 | |
3155 | l3->shared = new; | |
3156 | if (!cachep->nodelists[node]->alien) { | |
3157 | l3->alien = new_alien; | |
3158 | new_alien = NULL; | |
3159 | } | |
b28a02de PE |
3160 | l3->free_limit = (1 + nr_cpus_node(node)) * |
3161 | cachep->batchcount + cachep->num; | |
e498be7d CL |
3162 | spin_unlock_irq(&l3->list_lock); |
3163 | kfree(nc); | |
3164 | free_alien_cache(new_alien); | |
3165 | continue; | |
3166 | } | |
3167 | if (!(l3 = kmalloc_node(sizeof(struct kmem_list3), | |
b28a02de | 3168 | GFP_KERNEL, node))) |
e498be7d CL |
3169 | goto fail; |
3170 | ||
3171 | kmem_list3_init(l3); | |
3172 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
b28a02de | 3173 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; |
e498be7d CL |
3174 | l3->shared = new; |
3175 | l3->alien = new_alien; | |
b28a02de PE |
3176 | l3->free_limit = (1 + nr_cpus_node(node)) * |
3177 | cachep->batchcount + cachep->num; | |
e498be7d CL |
3178 | cachep->nodelists[node] = l3; |
3179 | } | |
3180 | return err; | |
b28a02de | 3181 | fail: |
e498be7d CL |
3182 | err = -ENOMEM; |
3183 | return err; | |
3184 | } | |
3185 | ||
1da177e4 LT |
3186 | struct ccupdate_struct { |
3187 | kmem_cache_t *cachep; | |
3188 | struct array_cache *new[NR_CPUS]; | |
3189 | }; | |
3190 | ||
3191 | static void do_ccupdate_local(void *info) | |
3192 | { | |
3193 | struct ccupdate_struct *new = (struct ccupdate_struct *)info; | |
3194 | struct array_cache *old; | |
3195 | ||
3196 | check_irq_off(); | |
3197 | old = ac_data(new->cachep); | |
e498be7d | 3198 | |
1da177e4 LT |
3199 | new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()]; |
3200 | new->new[smp_processor_id()] = old; | |
3201 | } | |
3202 | ||
1da177e4 | 3203 | static int do_tune_cpucache(kmem_cache_t *cachep, int limit, int batchcount, |
b28a02de | 3204 | int shared) |
1da177e4 LT |
3205 | { |
3206 | struct ccupdate_struct new; | |
e498be7d | 3207 | int i, err; |
1da177e4 | 3208 | |
b28a02de | 3209 | memset(&new.new, 0, sizeof(new.new)); |
e498be7d | 3210 | for_each_online_cpu(i) { |
b28a02de PE |
3211 | new.new[i] = |
3212 | alloc_arraycache(cpu_to_node(i), limit, batchcount); | |
e498be7d | 3213 | if (!new.new[i]) { |
b28a02de PE |
3214 | for (i--; i >= 0; i--) |
3215 | kfree(new.new[i]); | |
e498be7d | 3216 | return -ENOMEM; |
1da177e4 LT |
3217 | } |
3218 | } | |
3219 | new.cachep = cachep; | |
3220 | ||
3221 | smp_call_function_all_cpus(do_ccupdate_local, (void *)&new); | |
e498be7d | 3222 | |
1da177e4 LT |
3223 | check_irq_on(); |
3224 | spin_lock_irq(&cachep->spinlock); | |
3225 | cachep->batchcount = batchcount; | |
3226 | cachep->limit = limit; | |
e498be7d | 3227 | cachep->shared = shared; |
1da177e4 LT |
3228 | spin_unlock_irq(&cachep->spinlock); |
3229 | ||
e498be7d | 3230 | for_each_online_cpu(i) { |
1da177e4 LT |
3231 | struct array_cache *ccold = new.new[i]; |
3232 | if (!ccold) | |
3233 | continue; | |
e498be7d | 3234 | spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock); |
ff69416e | 3235 | free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i)); |
e498be7d | 3236 | spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock); |
1da177e4 LT |
3237 | kfree(ccold); |
3238 | } | |
1da177e4 | 3239 | |
e498be7d CL |
3240 | err = alloc_kmemlist(cachep); |
3241 | if (err) { | |
3242 | printk(KERN_ERR "alloc_kmemlist failed for %s, error %d.\n", | |
b28a02de | 3243 | cachep->name, -err); |
e498be7d | 3244 | BUG(); |
1da177e4 | 3245 | } |
1da177e4 LT |
3246 | return 0; |
3247 | } | |
3248 | ||
1da177e4 LT |
3249 | static void enable_cpucache(kmem_cache_t *cachep) |
3250 | { | |
3251 | int err; | |
3252 | int limit, shared; | |
3253 | ||
3254 | /* The head array serves three purposes: | |
3255 | * - create a LIFO ordering, i.e. return objects that are cache-warm | |
3256 | * - reduce the number of spinlock operations. | |
3257 | * - reduce the number of linked list operations on the slab and | |
3258 | * bufctl chains: array operations are cheaper. | |
3259 | * The numbers are guessed, we should auto-tune as described by | |
3260 | * Bonwick. | |
3261 | */ | |
3262 | if (cachep->objsize > 131072) | |
3263 | limit = 1; | |
3264 | else if (cachep->objsize > PAGE_SIZE) | |
3265 | limit = 8; | |
3266 | else if (cachep->objsize > 1024) | |
3267 | limit = 24; | |
3268 | else if (cachep->objsize > 256) | |
3269 | limit = 54; | |
3270 | else | |
3271 | limit = 120; | |
3272 | ||
3273 | /* Cpu bound tasks (e.g. network routing) can exhibit cpu bound | |
3274 | * allocation behaviour: Most allocs on one cpu, most free operations | |
3275 | * on another cpu. For these cases, an efficient object passing between | |
3276 | * cpus is necessary. This is provided by a shared array. The array | |
3277 | * replaces Bonwick's magazine layer. | |
3278 | * On uniprocessor, it's functionally equivalent (but less efficient) | |
3279 | * to a larger limit. Thus disabled by default. | |
3280 | */ | |
3281 | shared = 0; | |
3282 | #ifdef CONFIG_SMP | |
3283 | if (cachep->objsize <= PAGE_SIZE) | |
3284 | shared = 8; | |
3285 | #endif | |
3286 | ||
3287 | #if DEBUG | |
3288 | /* With debugging enabled, large batchcount lead to excessively | |
3289 | * long periods with disabled local interrupts. Limit the | |
3290 | * batchcount | |
3291 | */ | |
3292 | if (limit > 32) | |
3293 | limit = 32; | |
3294 | #endif | |
b28a02de | 3295 | err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared); |
1da177e4 LT |
3296 | if (err) |
3297 | printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n", | |
b28a02de | 3298 | cachep->name, -err); |
1da177e4 LT |
3299 | } |
3300 | ||
b28a02de PE |
3301 | static void drain_array_locked(kmem_cache_t *cachep, struct array_cache *ac, |
3302 | int force, int node) | |
1da177e4 LT |
3303 | { |
3304 | int tofree; | |
3305 | ||
e498be7d | 3306 | check_spinlock_acquired_node(cachep, node); |
1da177e4 LT |
3307 | if (ac->touched && !force) { |
3308 | ac->touched = 0; | |
3309 | } else if (ac->avail) { | |
b28a02de | 3310 | tofree = force ? ac->avail : (ac->limit + 4) / 5; |
1da177e4 | 3311 | if (tofree > ac->avail) { |
b28a02de | 3312 | tofree = (ac->avail + 1) / 2; |
1da177e4 | 3313 | } |
ff69416e | 3314 | free_block(cachep, ac->entry, tofree, node); |
1da177e4 | 3315 | ac->avail -= tofree; |
e498be7d | 3316 | memmove(ac->entry, &(ac->entry[tofree]), |
b28a02de | 3317 | sizeof(void *) * ac->avail); |
1da177e4 LT |
3318 | } |
3319 | } | |
3320 | ||
3321 | /** | |
3322 | * cache_reap - Reclaim memory from caches. | |
1e5d5331 | 3323 | * @unused: unused parameter |
1da177e4 LT |
3324 | * |
3325 | * Called from workqueue/eventd every few seconds. | |
3326 | * Purpose: | |
3327 | * - clear the per-cpu caches for this CPU. | |
3328 | * - return freeable pages to the main free memory pool. | |
3329 | * | |
3330 | * If we cannot acquire the cache chain semaphore then just give up - we'll | |
3331 | * try again on the next iteration. | |
3332 | */ | |
3333 | static void cache_reap(void *unused) | |
3334 | { | |
3335 | struct list_head *walk; | |
e498be7d | 3336 | struct kmem_list3 *l3; |
1da177e4 LT |
3337 | |
3338 | if (down_trylock(&cache_chain_sem)) { | |
3339 | /* Give up. Setup the next iteration. */ | |
b28a02de PE |
3340 | schedule_delayed_work(&__get_cpu_var(reap_work), |
3341 | REAPTIMEOUT_CPUC); | |
1da177e4 LT |
3342 | return; |
3343 | } | |
3344 | ||
3345 | list_for_each(walk, &cache_chain) { | |
3346 | kmem_cache_t *searchp; | |
b28a02de | 3347 | struct list_head *p; |
1da177e4 LT |
3348 | int tofree; |
3349 | struct slab *slabp; | |
3350 | ||
3351 | searchp = list_entry(walk, kmem_cache_t, next); | |
3352 | ||
3353 | if (searchp->flags & SLAB_NO_REAP) | |
3354 | goto next; | |
3355 | ||
3356 | check_irq_on(); | |
3357 | ||
e498be7d CL |
3358 | l3 = searchp->nodelists[numa_node_id()]; |
3359 | if (l3->alien) | |
3360 | drain_alien_cache(searchp, l3); | |
3361 | spin_lock_irq(&l3->list_lock); | |
1da177e4 | 3362 | |
e498be7d | 3363 | drain_array_locked(searchp, ac_data(searchp), 0, |
b28a02de | 3364 | numa_node_id()); |
1da177e4 | 3365 | |
e498be7d | 3366 | if (time_after(l3->next_reap, jiffies)) |
1da177e4 LT |
3367 | goto next_unlock; |
3368 | ||
e498be7d | 3369 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3; |
1da177e4 | 3370 | |
e498be7d CL |
3371 | if (l3->shared) |
3372 | drain_array_locked(searchp, l3->shared, 0, | |
b28a02de | 3373 | numa_node_id()); |
1da177e4 | 3374 | |
e498be7d CL |
3375 | if (l3->free_touched) { |
3376 | l3->free_touched = 0; | |
1da177e4 LT |
3377 | goto next_unlock; |
3378 | } | |
3379 | ||
b28a02de PE |
3380 | tofree = |
3381 | (l3->free_limit + 5 * searchp->num - | |
3382 | 1) / (5 * searchp->num); | |
1da177e4 | 3383 | do { |
e498be7d CL |
3384 | p = l3->slabs_free.next; |
3385 | if (p == &(l3->slabs_free)) | |
1da177e4 LT |
3386 | break; |
3387 | ||
3388 | slabp = list_entry(p, struct slab, list); | |
3389 | BUG_ON(slabp->inuse); | |
3390 | list_del(&slabp->list); | |
3391 | STATS_INC_REAPED(searchp); | |
3392 | ||
3393 | /* Safe to drop the lock. The slab is no longer | |
3394 | * linked to the cache. | |
3395 | * searchp cannot disappear, we hold | |
3396 | * cache_chain_lock | |
3397 | */ | |
e498be7d CL |
3398 | l3->free_objects -= searchp->num; |
3399 | spin_unlock_irq(&l3->list_lock); | |
1da177e4 | 3400 | slab_destroy(searchp, slabp); |
e498be7d | 3401 | spin_lock_irq(&l3->list_lock); |
b28a02de PE |
3402 | } while (--tofree > 0); |
3403 | next_unlock: | |
e498be7d | 3404 | spin_unlock_irq(&l3->list_lock); |
b28a02de | 3405 | next: |
1da177e4 LT |
3406 | cond_resched(); |
3407 | } | |
3408 | check_irq_on(); | |
3409 | up(&cache_chain_sem); | |
4ae7c039 | 3410 | drain_remote_pages(); |
1da177e4 | 3411 | /* Setup the next iteration */ |
cd61ef62 | 3412 | schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC); |
1da177e4 LT |
3413 | } |
3414 | ||
3415 | #ifdef CONFIG_PROC_FS | |
3416 | ||
85289f98 | 3417 | static void print_slabinfo_header(struct seq_file *m) |
1da177e4 | 3418 | { |
85289f98 PE |
3419 | /* |
3420 | * Output format version, so at least we can change it | |
3421 | * without _too_ many complaints. | |
3422 | */ | |
1da177e4 | 3423 | #if STATS |
85289f98 | 3424 | seq_puts(m, "slabinfo - version: 2.1 (statistics)\n"); |
1da177e4 | 3425 | #else |
85289f98 | 3426 | seq_puts(m, "slabinfo - version: 2.1\n"); |
1da177e4 | 3427 | #endif |
85289f98 PE |
3428 | seq_puts(m, "# name <active_objs> <num_objs> <objsize> " |
3429 | "<objperslab> <pagesperslab>"); | |
3430 | seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>"); | |
3431 | seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>"); | |
1da177e4 | 3432 | #if STATS |
85289f98 PE |
3433 | seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> " |
3434 | "<error> <maxfreeable> <nodeallocs> <remotefrees>"); | |
3435 | seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>"); | |
1da177e4 | 3436 | #endif |
85289f98 PE |
3437 | seq_putc(m, '\n'); |
3438 | } | |
3439 | ||
3440 | static void *s_start(struct seq_file *m, loff_t *pos) | |
3441 | { | |
3442 | loff_t n = *pos; | |
3443 | struct list_head *p; | |
3444 | ||
3445 | down(&cache_chain_sem); | |
3446 | if (!n) | |
3447 | print_slabinfo_header(m); | |
1da177e4 LT |
3448 | p = cache_chain.next; |
3449 | while (n--) { | |
3450 | p = p->next; | |
3451 | if (p == &cache_chain) | |
3452 | return NULL; | |
3453 | } | |
3454 | return list_entry(p, kmem_cache_t, next); | |
3455 | } | |
3456 | ||
3457 | static void *s_next(struct seq_file *m, void *p, loff_t *pos) | |
3458 | { | |
3459 | kmem_cache_t *cachep = p; | |
3460 | ++*pos; | |
3461 | return cachep->next.next == &cache_chain ? NULL | |
b28a02de | 3462 | : list_entry(cachep->next.next, kmem_cache_t, next); |
1da177e4 LT |
3463 | } |
3464 | ||
3465 | static void s_stop(struct seq_file *m, void *p) | |
3466 | { | |
3467 | up(&cache_chain_sem); | |
3468 | } | |
3469 | ||
3470 | static int s_show(struct seq_file *m, void *p) | |
3471 | { | |
3472 | kmem_cache_t *cachep = p; | |
3473 | struct list_head *q; | |
b28a02de PE |
3474 | struct slab *slabp; |
3475 | unsigned long active_objs; | |
3476 | unsigned long num_objs; | |
3477 | unsigned long active_slabs = 0; | |
3478 | unsigned long num_slabs, free_objects = 0, shared_avail = 0; | |
e498be7d | 3479 | const char *name; |
1da177e4 | 3480 | char *error = NULL; |
e498be7d CL |
3481 | int node; |
3482 | struct kmem_list3 *l3; | |
1da177e4 LT |
3483 | |
3484 | check_irq_on(); | |
3485 | spin_lock_irq(&cachep->spinlock); | |
3486 | active_objs = 0; | |
3487 | num_slabs = 0; | |
e498be7d CL |
3488 | for_each_online_node(node) { |
3489 | l3 = cachep->nodelists[node]; | |
3490 | if (!l3) | |
3491 | continue; | |
3492 | ||
3493 | spin_lock(&l3->list_lock); | |
3494 | ||
b28a02de | 3495 | list_for_each(q, &l3->slabs_full) { |
e498be7d CL |
3496 | slabp = list_entry(q, struct slab, list); |
3497 | if (slabp->inuse != cachep->num && !error) | |
3498 | error = "slabs_full accounting error"; | |
3499 | active_objs += cachep->num; | |
3500 | active_slabs++; | |
3501 | } | |
b28a02de | 3502 | list_for_each(q, &l3->slabs_partial) { |
e498be7d CL |
3503 | slabp = list_entry(q, struct slab, list); |
3504 | if (slabp->inuse == cachep->num && !error) | |
3505 | error = "slabs_partial inuse accounting error"; | |
3506 | if (!slabp->inuse && !error) | |
3507 | error = "slabs_partial/inuse accounting error"; | |
3508 | active_objs += slabp->inuse; | |
3509 | active_slabs++; | |
3510 | } | |
b28a02de | 3511 | list_for_each(q, &l3->slabs_free) { |
e498be7d CL |
3512 | slabp = list_entry(q, struct slab, list); |
3513 | if (slabp->inuse && !error) | |
3514 | error = "slabs_free/inuse accounting error"; | |
3515 | num_slabs++; | |
3516 | } | |
3517 | free_objects += l3->free_objects; | |
3518 | shared_avail += l3->shared->avail; | |
3519 | ||
3520 | spin_unlock(&l3->list_lock); | |
1da177e4 | 3521 | } |
b28a02de PE |
3522 | num_slabs += active_slabs; |
3523 | num_objs = num_slabs * cachep->num; | |
e498be7d | 3524 | if (num_objs - active_objs != free_objects && !error) |
1da177e4 LT |
3525 | error = "free_objects accounting error"; |
3526 | ||
b28a02de | 3527 | name = cachep->name; |
1da177e4 LT |
3528 | if (error) |
3529 | printk(KERN_ERR "slab: cache %s error: %s\n", name, error); | |
3530 | ||
3531 | seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", | |
b28a02de PE |
3532 | name, active_objs, num_objs, cachep->objsize, |
3533 | cachep->num, (1 << cachep->gfporder)); | |
1da177e4 | 3534 | seq_printf(m, " : tunables %4u %4u %4u", |
b28a02de | 3535 | cachep->limit, cachep->batchcount, cachep->shared); |
e498be7d | 3536 | seq_printf(m, " : slabdata %6lu %6lu %6lu", |
b28a02de | 3537 | active_slabs, num_slabs, shared_avail); |
1da177e4 | 3538 | #if STATS |
b28a02de | 3539 | { /* list3 stats */ |
1da177e4 LT |
3540 | unsigned long high = cachep->high_mark; |
3541 | unsigned long allocs = cachep->num_allocations; | |
3542 | unsigned long grown = cachep->grown; | |
3543 | unsigned long reaped = cachep->reaped; | |
3544 | unsigned long errors = cachep->errors; | |
3545 | unsigned long max_freeable = cachep->max_freeable; | |
1da177e4 | 3546 | unsigned long node_allocs = cachep->node_allocs; |
e498be7d | 3547 | unsigned long node_frees = cachep->node_frees; |
1da177e4 | 3548 | |
e498be7d | 3549 | seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \ |
b28a02de | 3550 | %4lu %4lu %4lu %4lu", allocs, high, grown, reaped, errors, max_freeable, node_allocs, node_frees); |
1da177e4 LT |
3551 | } |
3552 | /* cpu stats */ | |
3553 | { | |
3554 | unsigned long allochit = atomic_read(&cachep->allochit); | |
3555 | unsigned long allocmiss = atomic_read(&cachep->allocmiss); | |
3556 | unsigned long freehit = atomic_read(&cachep->freehit); | |
3557 | unsigned long freemiss = atomic_read(&cachep->freemiss); | |
3558 | ||
3559 | seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu", | |
b28a02de | 3560 | allochit, allocmiss, freehit, freemiss); |
1da177e4 LT |
3561 | } |
3562 | #endif | |
3563 | seq_putc(m, '\n'); | |
3564 | spin_unlock_irq(&cachep->spinlock); | |
3565 | return 0; | |
3566 | } | |
3567 | ||
3568 | /* | |
3569 | * slabinfo_op - iterator that generates /proc/slabinfo | |
3570 | * | |
3571 | * Output layout: | |
3572 | * cache-name | |
3573 | * num-active-objs | |
3574 | * total-objs | |
3575 | * object size | |
3576 | * num-active-slabs | |
3577 | * total-slabs | |
3578 | * num-pages-per-slab | |
3579 | * + further values on SMP and with statistics enabled | |
3580 | */ | |
3581 | ||
3582 | struct seq_operations slabinfo_op = { | |
b28a02de PE |
3583 | .start = s_start, |
3584 | .next = s_next, | |
3585 | .stop = s_stop, | |
3586 | .show = s_show, | |
1da177e4 LT |
3587 | }; |
3588 | ||
3589 | #define MAX_SLABINFO_WRITE 128 | |
3590 | /** | |
3591 | * slabinfo_write - Tuning for the slab allocator | |
3592 | * @file: unused | |
3593 | * @buffer: user buffer | |
3594 | * @count: data length | |
3595 | * @ppos: unused | |
3596 | */ | |
b28a02de PE |
3597 | ssize_t slabinfo_write(struct file *file, const char __user * buffer, |
3598 | size_t count, loff_t *ppos) | |
1da177e4 | 3599 | { |
b28a02de | 3600 | char kbuf[MAX_SLABINFO_WRITE + 1], *tmp; |
1da177e4 LT |
3601 | int limit, batchcount, shared, res; |
3602 | struct list_head *p; | |
b28a02de | 3603 | |
1da177e4 LT |
3604 | if (count > MAX_SLABINFO_WRITE) |
3605 | return -EINVAL; | |
3606 | if (copy_from_user(&kbuf, buffer, count)) | |
3607 | return -EFAULT; | |
b28a02de | 3608 | kbuf[MAX_SLABINFO_WRITE] = '\0'; |
1da177e4 LT |
3609 | |
3610 | tmp = strchr(kbuf, ' '); | |
3611 | if (!tmp) | |
3612 | return -EINVAL; | |
3613 | *tmp = '\0'; | |
3614 | tmp++; | |
3615 | if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3) | |
3616 | return -EINVAL; | |
3617 | ||
3618 | /* Find the cache in the chain of caches. */ | |
3619 | down(&cache_chain_sem); | |
3620 | res = -EINVAL; | |
b28a02de | 3621 | list_for_each(p, &cache_chain) { |
1da177e4 LT |
3622 | kmem_cache_t *cachep = list_entry(p, kmem_cache_t, next); |
3623 | ||
3624 | if (!strcmp(cachep->name, kbuf)) { | |
3625 | if (limit < 1 || | |
3626 | batchcount < 1 || | |
b28a02de | 3627 | batchcount > limit || shared < 0) { |
e498be7d | 3628 | res = 0; |
1da177e4 | 3629 | } else { |
e498be7d | 3630 | res = do_tune_cpucache(cachep, limit, |
b28a02de | 3631 | batchcount, shared); |
1da177e4 LT |
3632 | } |
3633 | break; | |
3634 | } | |
3635 | } | |
3636 | up(&cache_chain_sem); | |
3637 | if (res >= 0) | |
3638 | res = count; | |
3639 | return res; | |
3640 | } | |
3641 | #endif | |
3642 | ||
00e145b6 MS |
3643 | /** |
3644 | * ksize - get the actual amount of memory allocated for a given object | |
3645 | * @objp: Pointer to the object | |
3646 | * | |
3647 | * kmalloc may internally round up allocations and return more memory | |
3648 | * than requested. ksize() can be used to determine the actual amount of | |
3649 | * memory allocated. The caller may use this additional memory, even though | |
3650 | * a smaller amount of memory was initially specified with the kmalloc call. | |
3651 | * The caller must guarantee that objp points to a valid object previously | |
3652 | * allocated with either kmalloc() or kmem_cache_alloc(). The object | |
3653 | * must not be freed during the duration of the call. | |
3654 | */ | |
1da177e4 LT |
3655 | unsigned int ksize(const void *objp) |
3656 | { | |
00e145b6 MS |
3657 | if (unlikely(objp == NULL)) |
3658 | return 0; | |
1da177e4 | 3659 | |
065d41cb | 3660 | return obj_reallen(page_get_cache(virt_to_page(objp))); |
1da177e4 | 3661 | } |
543537bd PM |
3662 | |
3663 | ||
3664 | /* | |
3665 | * kstrdup - allocate space for and copy an existing string | |
3666 | * | |
3667 | * @s: the string to duplicate | |
3668 | * @gfp: the GFP mask used in the kmalloc() call when allocating memory | |
3669 | */ | |
dd0fc66f | 3670 | char *kstrdup(const char *s, gfp_t gfp) |
543537bd PM |
3671 | { |
3672 | size_t len; | |
3673 | char *buf; | |
3674 | ||
3675 | if (!s) | |
3676 | return NULL; | |
3677 | ||
3678 | len = strlen(s) + 1; | |
3679 | buf = kmalloc(len, gfp); | |
3680 | if (buf) | |
3681 | memcpy(buf, s, len); | |
3682 | return buf; | |
3683 | } | |
3684 | EXPORT_SYMBOL(kstrdup); |